Changelog
What changed in FlockTab
Every change you can see, newest first: the tab CLI, the console, the proxy, Agent Guard and billing.
Follow it in a feed reader: /changelog.xml. Update tab with tab update; it shows the same notes.
tab CLI 0.1.87
- Pooled Claude logins keep renewing: no lock is linked from ~/.claude.
- A lock link that an older tab made goes at the next pooled launch.
Proxy
- Fixed: Codex on a ChatGPT subscription no longer stalls on the hosted proxy. The proxy refused Codex's account calls with a 401, which Codex takes as its login expiring, so it signed in again several times a second and never sent the turn. It now answers 403.
tab CLI 0.1.86
- The pool moves a running session only between turns, never mid-task.
- A move waits for background jobs. Each move goes to pool-moves.log.
tab CLI 0.1.85
- tab codex send starts a turn in an idle session; no more stuck queues.
- A message Codex may have taken is never queued again, so it runs once.
- A pooled run now removes its run records when it ends.
tab CLI 0.1.84
- Self-hosted on subscriptions only: no API keys asked, the proxy starts.
- An API-key Agent there is told which key is missing; tab up adds it.
- The pool splits sessions across logins by the room each one has.
Console
- Fixed: the subscription pool now splits sessions across your logins by the room each one has, instead of sending them all to the emptiest. Sessions started one after another on one machine count at once, and a login added to the pool is used by sessions that were already running. Needs the next tab release.
Proxy
- Fixed: tab codex on a ChatGPT subscription starts again on the hosted proxy. Codex 0.162 checks your account when it starts, and the proxy refused that check with a 401.
- Fixed: a self-hosted machine with only subscriptions no longer asks for API keys. The self-hosted proxy now starts with no provider keys, and your Claude, ChatGPT, Grok or Kimi logins pass through as before. An API-key Agent on that machine is told which key is missing, and tab up adds it. Needs the next tab and proxy release.
tab CLI 0.1.83
- tab pages: report pages with a comment box; comments reach a session.
- tab agy: Antigravity joins the flock, gets messages and steers Codex.
- tab codex send steers a busy turn; tab flock shows what each is doing.
Console
- New: tab pages shows your agents' report pages on your own machine, with a comment box on every item. A comment can go to a named agent at once, and the agent's reply shows on the page as soon as it is written. Needs the next tab release.
Console
- tab codex send now puts a message into a busy Codex session's current turn, without waiting for the turn to end. This works for sessions started with plain tab codex after this update. Others still queue the message. Needs the next tab release.
- Fixed: when tab is closed by force, its Codex server now stops at the next tab codex send, and never while its Codex window still runs. Needs the next tab release.
- A message from a named Antigravity session to a busy Codex session now goes into its current turn too, as one from Claude Code or Codex does. Needs the next tab release.
- tab flock and the send commands now say what a session is doing and how many messages wait for it. A message behind a long turn no longer looks lost. Needs the next tab release.
- tab flock now also says how many messages wait for an Antigravity session. Needs the next tab release.
- Every short session id that tab prints now names one session only, even for two sessions started in the same minute. This covers Claude Code, Codex and Antigravity sessions. Needs the next tab release.
- tab agy now starts Antigravity with its own copy of your settings, so it runs through tab. Running agy on its own leaves your settings unchanged. Needs the next tab release.
- Antigravity started with tab agy can now name its session, list the flock, and message Claude Code and Codex sessions by name. Needs the next tab release.
- You can now message an Antigravity session by name with tab agy send, or ask Claude Code or Codex to. A busy session reads it before its next step, and an idle one when you next type. Needs the next tab release.
- Fixed: a message to an Antigravity session is never left stuck after a crash. Needs the next tab release.
tab CLI 0.1.82
- Claude Code in tab claude messages Codex by name when you ask in words.
- Claude Code can also name its session and list the flock when asked.
tab CLI 0.1.81
- tab flock names Claude and Codex sessions; Codex can message Claude too.
- tab claude send, tab codex send: message a named session, get a reply.
- tab claude compacts a quiet session once its context passes 400k.
tab CLI 0.1.80
- tab top: each login's week at its 6-hour pace, and if the pool lasts.
- Pool: tab pool spread, pause until a reset; running Agents see a pause.
- tab top --streamer: made-up names for logins, emails, Agents, projects.
Agent Guard
- The loop watch now catches a Claude Code agent that gives the same reply three times in a row with no tool call, such as a goal check that keeps saying continue. Needs the next tab release.
tab CLI 0.1.79
- Pool: with Agent Guard, busy sessions leave full logins mid-reply.
- Pool: a background job delays a needed move by 10 minutes at most.
Console
- Fixed: an agent that never pauses between replies now leaves a login over its limit too. Ten minutes after it had to move, with Agent Guard on, tab stops its next tool call with a note to run it again, and moves it to another login in the same conversation. Before, it stayed until the login was used up. Needs the next tab release.
- Fixed: background jobs hold a move off a login over its limit for 10 minutes at most. A Monitor that re-arms, or a loop that never ends, used to keep a session there until the login ran out. Needs the next tab release.
- Fixed: tab pool now counts every session on a login when it shares out logins, not one per Agent. Many sessions in one folder used to count as one, so new sessions kept landing on the same login. Needs the next tab release.
tab CLI 0.1.78
- Agent Guard: allow a call it held for a secret once, from the console.
- Agent Guard: fewer false holds on names, links and your own questions.
- tab top: draws at once, groups subscriptions by harness, shows ctx.
Agent Guard
- Turning Agent Guard off for an Agent now applies to its running sessions at once, not only from their next launch. Allowing a held call with tab guard allow now opens the console's confirm page for you. Needs the next tab release.
- A message to another agent with a link to claude.ai, github.com, flocktab.com or figma.com is no longer held for the page id in the link. A value after ? or # in a link is still checked. Needs the next tab release.
Agent Guard
- You can now let one call through that Agent Guard refused because it carries a secret, such as a message to another agent it misread. The call is held like any other, with the secret hidden. Run tab guard allow on that machine and confirm it in the console: exactly that call runs once. It can never be allowed always, and the console only gets a fingerprint keyed to your machine, never the text. Needs the next tab release.
- Agent Guard no longer judges a question your agent asks you (Claude Code's AskUserQuestion), since your answer is the decision. A question that would show a real secret value stays refused. Needs the next tab release.
- Fixed: names that run words and short numbers together no longer get a message to another agent refused by Agent Guard, such as dash-first4figures-com, holeDownTrue or a 1440x900 screenshot. A random-looking id, such as a link's, is held as before. Needs the next tab release.
- The loop watch now says "Waiting on a timer" for an agent that only re-runs one read-only check every minute or so, such as listing its open tasks, instead of "Stuck: repeating itself". It still counts as idle, as before.
- Fixed: Agent Guard no longer refuses a status message to another agent that names a secret and says what happened to it, such as "SENTRY_AUTH_TOKEN: removed from the guard". A value after the name is still refused, however short, as is anything in quotes.
- Fixed: Agent Guard no longer refuses a python script that only edits files on this machine, run as python3 - with a quoted here-document, or a helm template render. A script that imports anything beyond re, json, pathlib and textwrap, or can run or load other code, is still checked as before, and so are helm install, upgrade and remote charts.
- Fixed: text that cat or tee writes to a file is no longer read as an agent approving its own held call, such as an audit note that quotes the review endpoint. Fed to a shell, or beside any program that is not known to stay on this machine, it is still held.
- Agent Guard now checks the messages your agents send to other agents and AIs: a Claude Code message to another session, a Codex hand-off, the Codex MCP server, and a shell command that calls an AI provider's API. A message that carries a key, a token, a password or any long random-looking string is refused on your machine before it leaves, test keys included, in every mode but off; a full commit SHA from your own repository passes, and a message the guard cannot read is refused. A long hand-off, up to 1 MiB, is checked rather than refused. The refusal names the kind and who it was for, never the value, and the message itself is not kept. In Claude Code, a reply from the Codex MCP server reaches your agent marked as data, never as instructions. Needs the next tab release.
- Fixed: Agent Guard no longer refuses a note or memory that describes the shape of a credential with a placeholder in angle brackets, such as password=(<16-char value>). A real value in the same place is still refused. Needs the next tab release.
- Fixed: Agent Guard no longer refuses a local search because of words in it. Text in quotes or in a quoted here-document is data, never a command, and git grep, git show of a file, if, while and for loops and $( ) of local readers stay on your machine. A command that can send, such as curl, ssh or a git push, is still refused when it carries a secret. Needs the next tab release.
- Fixed in the Agent Guard docs: a SendMessage is checked even between a subagent and its lead, since its address cannot tell a teammate from another session. The docs said a session's own subagents were never checked; only those started with the Agent tool are not.
Console
- tab top opens at once: it shows the screen from your last run, marked with how old it is, until the first read lands a moment later, instead of a blank wait while every panel was read. q now quits straight away, even in the middle of a read. Needs the next tab release.
- Fixed: tab top's ctx for a Claude Code session on this machine is now what that session's own status line says. Before, it showed the session's largest recent call, which can be a subagent's or a compaction's: 815k and /compact now for a conversation at 326k. Sessions on other machines still show their largest call. Needs the next tab release.
Site and docs
- New Compare page: FlockTab beside the spend limits of OpenAI, Anthropic and Google and five AI gateways (LiteLLM, Portkey, Helicone, Cloudflare AI Gateway, OpenRouter), what each does well, and every claim linked to that vendor's own docs.
- The line under the supported harnesses table now says in plain words when the live runs were done, instead of naming tab releases.
tab CLI 0.1.77
- tab top: [compact] runs /compact in an Agent's sessions due for one.
tab CLI 0.1.76
- OpenCode, Cline and Pi: Agent Guard refuses calls if its Node is gone.
tab CLI 0.1.75
- Agent Guard keeps working after brew upgrades Node; no unchecked calls.
- Codex, Grok and Kimi follow your global AGENTS.md, as Claude does.
- tab top: [pause] and [next] per pooled login; tab pool next <name>.
tab CLI 0.1.74
- Run tab plugin update once; from now on tab update does it for you.
- Without a terminal (a pipe, a script, CI), tab no longer starts a login.
- tab list and tab status say who closed a tab; the loop watch says why.
Agent Guard
- Fixed: under OpenCode, Cline and Pi, Agent Guard now refuses a call when its own Node or script is missing, with "FlockTab Agent Guard is missing: relaunch with tab, or run tab setup", as it already does under Claude Code, Codex and the others. Before, those three let the call run unchecked. Needs the next tab release.
- Fixed: after Homebrew upgraded Node, Agent Guard could stop checking calls without a word, because tab had written the exact Node version's path, which the upgrade deleted. tab now writes the path Homebrew (or mise) keeps across upgrades, and if the guard's Node or script is ever missing, the call is refused with "FlockTab Agent Guard is missing: relaunch with tab, or run tab setup" instead of running unchecked. Needs the next tab release.
- The Agent Guard card on Overview and the guard line in tab top count stuck against the loop watch's own checks (for example, loop 41/380 stuck), not beside Agent Guard's tool checks, where it read as a share of them.
Console
- tab top now has a [compact] button beside the context of an Agent that has a Claude Code session on this machine due for a compact. One click runs /compact in each of those sessions, as if you typed it, once that session is idle; the button reads compacting until it has run. Only sessions started with tab claude by this tab or newer get it. Needs the next tab release.
- When a vendor trims a subscription login's limit, FlockTab says so: Subscriptions marks the login, every console page shows a notice, tab top tags it (5h limit ↓22%), and Alerts can send it to Slack, Discord, Teams or a webhook. It is estimated from the login's own calls, and a change of plan or of our prices is never taken for a trim.
- tab top now shows how full each session's context is, beside its Agent: every session's figure (850k 412k 90k), amber when it should compact soon and red when it should /compact now, the same limits as the status line, with how many sessions need one next to the Agents title. It is what each call sends, cached input included, read from the calls FlockTab already records. tab top assumes Claude Code's 1M window, so a Codex, Grok or Kimi session shows its figure without colour. A tab before this release leaves the column out.
- Codex under tab, and Codex, Grok and Kimi logins in the tab pool, now follow your own global instructions: tab links your ~/.codex/AGENTS.md (or the AGENTS.md in ~/.grok and ~/.kimi-code) into the folder it runs them from, as Claude already had with CLAUDE.md. Before, they ran with none of your rules. It happens at the next launch; nothing to run. Needs the next tab release.
- Fixed: running tab login before you have an account no longer loses the code. Start one from the login page, and after signing up you land on the page that approves this machine, with its code filled in. Before, you landed on the console and had to find that page and type the code. The setup screen also tells people on Claude Max or ChatGPT Plus to choose Subscription when tab asks which kind of Agent.
Proxy
- The self-hosted proxy answers --version and --help, and refuses an option it does not know. Before, any of them started a proxy with your keys.
Site and docs
- Fixed: on a phone, the diagram of a call held against a tab (on How it works, Product and the home page) now stacks its three steps, so the CLOSED mark and "$48.18 of $50.00" no longer spill out of their box. The with and without FlockTab demo on How it works stacks its two terminals on a phone too, so you can read both logs, and its red BLOCKED lines are easier to read on the dark background.
- Fixed: with reduced motion turned on, the strip of providers on the home page no longer stops cut off at the edge. It shows every provider once, wrapping onto more lines.
- Fixed: on a phone, the "How Agent Guard works" link on the home page, the links to the other features on each feature page and the See also links in the tab CLI docs are now at least 24 px tall, so they are easier to tap.
- tab and the self-hosted proxy are proprietary software of Hanamori Labs, LLC, used under the FlockTab terms. From their next release, their npm packages say so (license UNLICENSED, with a LICENSE file) instead of MIT. Nothing changes in how you install or use them.
tab CLI 0.1.73
- When FlockTab can't be reached, tab says which part is down, plainly.
- tab up never asks a machine on subscriptions only for provider keys.
- tab keeps your Agent's key out of the addresses its harness calls.
tab CLI 0.1.72
- Agent Guard: a stuck check fails in time, not let past (Mac, Linux).
tab CLI 0.1.71
- tab top: a long branch name is cut with …, so its columns line up.
- Agent Guard: long calls are checked fast; over 1 MB is a failed check.
tab CLI 0.1.70
- tab top: each Agent's CPU, memory and processes on this Mac.
tab CLI 0.1.69
- Agent Guard: a word full of * no longer lets a call past unchecked.
tab CLI 0.1.68
- Band: the gap counts wrapped rows, so a long line is never clipped.
tab CLI 0.1.67
- Band: one blank row above it when it has room; squeezed, rows win.
tab CLI 0.1.66
- Your own status line is back under the prompt in tab claude.
- Band: no Session row; the session's numbers stay in your status line.
tab CLI 0.1.65
- Band: FlockTab shows once, in the band or the status line, never none.
- Band: a folder on no tab says so: not on a tab · tab use claude.
- Pool: logins' keychain dates are read every 15 min, not each minute.
tab CLI 0.1.64
- Band: a Session row shows model, context, cost, lines and branch.
- Band only: in tab claude your own status line gives way to the band.
- Agent Guard: on its own error it asks or denies; reads still run.
tab CLI 0.1.63
- Band: FlockTab's row puts the numbers first and the rest dim.
- Band: the notifications tip shows only when a number is at 80%+.
Agent Guard
- When the loop watch closes a stuck Agent's tab, the tab now says so, with the check's verdict, next to Reopen. Before, it looked like a close by hand.
- On macOS and Linux, a stuck Agent Guard check can no longer let a call through. Harnesses such as Claude Code and Gemini CLI run a call once its guard has taken too long, so a small watchdog now gives the "ran out of time" answer in its place first: asked in Ask first, refused in Deny. Cline, which waits under 3 seconds, now gets its answer in time too. Windows has no watchdog; there, the check's own time limit is the bound. Needs the next tab release.
- Agent Guard now reads very long commands quickly: some of its rules took seconds on 100 KB, long enough for the harness to stop waiting and run the call unchecked. If a check still runs out of time, it stops and answers as a failed check (asked in Ask first, refused in Deny), and a tool call over 1 MB is answered that way without being checked. Plain reads still run, and Watch lets everything through. Needs the next tab release.
- Fixed: a command with a long run of * in one word, such as cat ~/.claude/ followed by 16 stars, no longer keeps Agent Guard busy for many seconds. On the machine, a check that slow ran past the harness's hook time limit, and the harness then ran the call with no check. Such a word is now checked in a moment, with the same result as before. Needs the next tab release.
Console
- tab no longer puts your Agent's key in the addresses its harness calls: each launch gets its own short-lived link instead, and Claude Code and Codex send the key separately, so an address seen in a log is no use to anyone. Needs the next tab release.
- The drift watch cards on Models now also show, for each model, how many answers the vendor refused or filtered, how many were cut short by a length limit, and how many thinking tokens a call used, against the model's usual day. FlockTab reads why each reply ended and its thinking count from the vendor's own reply, from now on; no reply text.
- New on Models: a drift watch card for each of your busiest models. It sets the latest day against the model's usual day (time to the last byte, output tokens and cost per call, calls the vendor failed) and marks a change of a quarter or more. It also checks the model the vendor's reply names against the one each call asked for, and lists any other model in red. Read from what every call already records; no prompt or reply text, and no extra calls.
- Fixed: a Vercel connection that Vercel answers with "not found" for any reason other than "no costs" now shows that on its card, instead of reading as $0. Only Vercel's own "no costs" answer counts as a $0 bill.
- Fixed: Vercel spend syncs again for busy teams. Since October 1 the sync stopped with "response too large" and Outside spend showed Vercel at $0. Vercel's billing export has no pages, so a month that is too large for one answer is now read a few days at a time; the lines already stored stay, with no duplicates.
- Each connected outside source now has its own page. Open it from the source's name or View details on its card under Outside spend. It shows the source's accounts and their last sync, spend per day by service, and what it cost by service, resource, project, account and month, each against the period before, with a month-end estimate. Click a row or a day to see its lines, and export that source alone as CSV. A failed sync now says in plain words what happened and what to do. Railway and DigitalOcean send no daily figures, so their pages say so, and DigitalOcean lists its invoices.
Proxy
- The self-hosted proxy now also tells flocktab.com the model the vendor's reply named, why the reply ended and its thinking tokens, so calls through it count on the Models drift watch cards too. Names and words only, no reply text. Fixed too: until now the self-hosted proxy read no usage from streamed Gemini replies on subscription and WebSocket calls (Google separates stream events differently), so those past self-hosted Gemini calls may show as unpriced. Needs the next proxy release.
tab CLI 0.1.62
- Band only: FlockTab's line leaves the status line where the band draws.
- Band: FlockTab's line comes first, so a short band keeps it.
- Band plugin: Claude Code before 2.1.287 no longer refuses it.
tab CLI 0.1.61
- Band: with no room above the prompt, the status line keeps the line.
- Band: FlockTab's line shows a second after start, not five.
tab CLI 0.1.60
- Pool: a moved session says which login it's on and to re-arm its jobs.
- Pool: logins near the end of their week get new sessions last.
tab CLI 0.1.59
- FlockTab's line sits above Claude Code's prompt (Claude Code 2.1.287+).
- tab setup: login, Agent, alias, status line, plugin in one check list.
- Pool: on a full login, a session moves once its reply and jobs end.
Agent Guard
- When Agent Guard's own check fails on a call, the call no longer runs unchecked. In Ask first you are asked, and in Deny it is refused; plain reads of the machine still run, and Watch lets everything through as before. The failed call is listed on the Agent with the error's name, so you can see it happened. Switch an Agent to Watch to let it work while a fix ships. Needs the next tab release.
- Fixed: Agent Guard no longer fails on a command with a [ that never closes, such as a grep pattern like [^ next to a path. Such a command made the check error out, and the tool call then ran without any rule checking it. Agent Guard now reads the word the way the shell does, as a plain name.
- Fixed: the console no longer shows the red "Judgments unavailable, rules only" strip because a single Agent Guard judgment was slow. At a quiet time, one answer that took too long counted as an outage. It now takes at least three checks on the rules alone, and more than one in ten, before the strip appears.
Agent Guard
- Fixed: Agent Guard no longer reads a variable named for a secret as the secret. A line of code such as token: latestToken or password: new_password, in an edit an agent makes with a script, is a name and not a value. A single word, a value with a digit or a capital first letter, and anything in quotes still count. Needs the next tab release.
- Fixed: Agent Guard no longer stops an agent from starting a local test database in Docker. docker run of a stock database image (Postgres, MySQL, MariaDB, MongoDB, Redis) with its password as a setting stays on your machine, so it passes when Docker is using this machine's own daemon. A remote Docker host, any other image, a command run inside the container, docker push and docker login are still stopped. Needs the next tab release.
- Fixed: Agent Guard no longer reads ordinary code as a password or token. Code that fetches a secret (token: user.token, password: params[:password], secret: Config.fetch(:secret)) is not the secret, in a command, a question or a push. In a pushed source file, an unquoted name is a variable whatever it is called (token: other_token), and a search pattern for where tokens are used (grep '\$[A-Z_]*TOKEN') is not a token. A value still stops the call: anything quoted, a bare password in a command, and in a pushed file a value in a comment, inside a string, or in YAML, a shell script or notes. With the next tab.
- Fixed: Agent Guard no longer stops two things that only looked like secrets. A git push whose commits pass a variable on under its own name (super_token: super_token) is not a password or token leaving, and a page address with a port and an @ later in it (http://127.0.0.1:5295/page?poses=idle@0.5) is not a password in a web address. A real password or token in a pushed line, in a remote's address or given to git on the command line is refused as before. With the next tab.
Proxy
- Fixed: a Claude subscription that has run out now shows as full in tab top and the console, with the time it comes back, instead of staying at its last reading (90%, say) while Anthropic refuses every call. The pool no longer picks a login that is out. With the self-hosted proxy this needs the next proxy release.
Site and docs
- Fixed: flocktab.com's pages load 44% less CSS (about 87 KB instead of 157 KB, in one file instead of two), so they show sooner, most of all on phones. The site was shipping styles for code its pages never use.
- Faster flocktab.com, most of all on phones: every page now downloads a quarter less JavaScript, because the public site no longer ships an error-reporting library it never used.
tab CLI 0.1.58
- tab top reads logins, pool and Agent Guard every 10 s, not every draw.
- Outside spend and branches refresh every 30 s; press r to read all now.
- The pool reuses the logins tab top just read: one console call, not two.
Agent Guard
- Fixed: Review this command on a tip opens that command on the Agent Guard page, with Always allow ready, instead of the whole page. The tip's other links also go straight to the right page.
- Fixed: the tip about a command you allowed by hand again and again now shows that command and the Agent that ran it, and its button opens that exact command on the Agent Guard page with Always allow ready (no longer the whole page). Allowing it always there clears the tip; if the command is already allowed or too old, the page says so.
Console
- A machine can now end its own tab login on FlockTab's side, not just forget it locally: with the next tab, tab logout stops that machine's session at once, and every proxy address it handed out stops working with it. Other machines logged in to the same flock stay signed in.
- Fixed: closing your flock now also removes the apps you connected to it over MCP and any single sign-on that was still in progress, so nothing is left that names the closed flock.
- You can close your flock from Settings. The owner types the flock's name to confirm; the console first says what goes, that a paid plan's subscription ends right away, and who else in the flock loses access. Closing cancels the subscription, erases the flock and everything in it, and signs you out. If the subscription can't be cancelled, nothing is erased.
Site and docs
- The Privacy Policy now also names Google Analytics and the TikTok and Meta pixels on flocktab.com's public pages, alongside X, with the same rules: nothing loads in the EEA, the UK or Switzerland until you accept, Global Privacy Control turns the ad pixels off, and none of them run on the console.
- flocktab.com's public pages now measure our ads with X's advertising pixel, loaded through Google Tag Manager, and never on the console or the sign-up, sign-in and checkout pages. In the EEA, the UK and Switzerland a banner asks first, with Accept and Reject side by side, and nothing loads until you accept. Elsewhere, Your privacy choices at the bottom of every page turns it off, and a browser that sends Global Privacy Control is opted out automatically. Cookie settings changes your answer at any time. The Privacy Policy now describes this in a new section, Advertising on our public pages.
- Fixed: faster home page, most of all on phones. The 3D pictures now start only on a computer with a graphics card, once the page has loaded and as you scroll to them; phones and slower machines keep the still pictures. flocktab.com now also always uses HTTPS and can't be shown inside another site, and screen readers hear the list of providers it works with.
- New Videos page (flocktab.com/videos): every FlockTab film in one place, from the 45-second overview and how it works to the tab CLI, Agent Guard and ten 20-second shorts, all narrated in the same new voice with captions on. Nothing downloads until you press play, and each film can be saved in 16:9 or 9:16 to share.
- Sharing your leaderboard page now says something worth clicking: the post leads with your flock's best figure from the last 7 days and invites people to start using FlockTab, the card carries a "Start using FlockTab today" line and always says when its figures were taken, and visitors to your page see a Start your own tab button at the top. A sign-up that comes through your share still credits you, and now also records which network it came from.
- The Terms and Privacy Policy now say exactly what closing a flock does: its subscription ends at once with no refund for the rest of the period (an ordinary cancellation still runs to the period's end), and everything in the flock is deleted, while support tickets, your own sign-in and security settings, and billing records are kept.
tab CLI 0.1.57
- Fixed: Agent Guard mistook test fixtures and localhost for secrets.
- The status line says run tab login when this machine's login lapses.
- tab restore brings an archived Agent back; after 24 h it needs a seat.
Agent Guard
- Agent Guard grows with your flock: each Agent adds 5,000 Agent Guard checks a month, shared across your flock, up to 500,000 (Team had 20,000 in all; Fleet never has fewer than its 200,000). An Agent counts while it holds a seat. The Plan page shows how many checks are used this month, and the Agent Guard page and the notice when they run low or out give the same figure, with a link to add an Agent for 5,000 more. Solo has no Agent Guard: it comes with Team.
- Agent Guard speaks as Agent Guard everywhere: the monthly cap notice, the Agent Guard page's count (Judgments this month), the guard levels, the notice when judgments are unavailable, tab help and the privacy policy no longer use the name of the model behind it.
- Fixed: the secrets block no longer refuses local work. A password in an address on your own machine (a localhost database URL) never counts, nor does a placeholder or a reference such as ${env:API_KEY}, $API_KEY, %API_KEY%, <token> or PASTE_YOUR_KEY, and searching for a key with grep, rg, sed, awk, jq or find is fine, also behind rtk, env, nice or timeout. Nor does a test fixture's value (public-fixture-key) or an unsigned token from the Firebase Auth emulator, and a browser tool (Claude in Chrome, Playwright, Chrome DevTools) may open a page on your own machine with a token in its address. Sending one with curl, wget, a script, a push, or a browser to another host is still refused. Needs the next tab release.
- Fixed: calls Agent Guard refused on its own, such as a secret on its way off your machine, now show under Blocked in Agent Guard's Review, not only under All.
- Agent Guard refuses a tool call that would send a secret off your machine, at once and without asking: a live API key, token or password in a curl or wget request, a web fetch, an email, Slack or MCP tool, or a git push that adds an .env file, a key file or a line holding a key. The refusal says what kind of secret and where it was going, never the value. Writing a secret to a local file such as .env is fine, and test keys and placeholders never count. Needs the next tab release.
Billing
- Fixed: a Claude Code window that moved to another pool login could end with "No conversation found" when you had switched conversations inside it with /resume. The move now reopens the conversation the window was actually in. Needs the next tab release.
- In Claude Code, tab's status line says when to compact: ctx 517k · compact soon in amber from 400k tokens per call (the same line as the Overview tip) or a window 60% full, and /compact now in red only when the window is 80% full, just before Claude Code compacts on its own. Needs the next tab release.
Console
- Fixed: closing your flock now works whatever it holds. A flock that had used subscription logins, outside spend entries or alerts, two-factor or passkeys, or members, roles, SSO or SCIM could not be closed; now those go with it.
- tab top asks the console for less: who is working and the Agents on every redraw, the logins, the pool and Agent Guard every 10 seconds, outside spend and branches every 30, and r still reads everything at once. The console answers every tab call a little faster.
- Fixed: the counts no longer contradict each other. "Agents" always means the Agents that run, and seats are shown on their own: the sidebar reads, for example, 21 agents · 22/22 seats (1 frees 1 Oct 14:32), since an Agent archived in the last 24 hours keeps its seat until then. The same words are on the Overview and the Plan page, and the Agents count in the menu is the live Agents only. When every seat is in use, New tab says so in seats, with the time the next one frees, and the console gives that time on your flock's own clock.
- Search (⌘K) now finds where the money goes: type Cloudflare or Expo and the source comes first with its last 30 days, then budgets, projects, logins, your plan and invoices, and Agent Guard rules, each opening its page. Ledger results say which field matched, and a call no longer matches on a tool it only declared.
- Fixed: a tip's button opens the page of what it is about. Model spend growing on OpenRouter opens OpenRouter (any other source opens its own card), a login's tip opens that vendor's logins, and a command allowed by hand opens Agent Guard's held commands; a source that stopped syncing always opens its own card. Cards on Outside spend › OpenRouter, Settings, Security and Enterprise span the full row instead of leaving one card at half width beside empty space.
- A machine's tab login now keeps who approved it, and every call it makes checks that person can still manage Agents and keys for the whole flock. Narrow their role, suspend or remove them, and the machine stops at its next call until it logs in again. A login approved before today keeps working on a flock with no other members; once the flock has its first active member, machines logged in before today have to run tab login again.
- Log out now ends that session on FlockTab's side too: a copy of the browser's session taken before you logged out stops working at once, instead of lasting until it expired (up to an hour). Your sessions on other devices stay signed in.
- Fixed: the ticker across the top kept an Agent you had just archived until it refreshed on its own. It now drops it straight away.
- Fixed: Archive agent archived on the first click, and Close all tabs could close every tab on one click. The first click now only arms the button (it reads Click again to archive, or Click again to close all, with a bar showing the time left); only a second click does it. Escape or waiting cancels.
- Archive agent is on the Agents page, and replaces Delete tab on a tab's page. Before you confirm, it says what happens: the Agent stops at once, its calls are refused, its history stays, and its seat frees 24 hours later, at the time shown. Until then an archived Agent still counts towards your plan's Agents and the Agents Team pays for, so a new Agent in its place waits, and the Plan page will not lower your Agents below the ones holding a seat.
- Restore an archived Agent: the Agents page lists your archived Agents, when each was archived and when its seat frees, with a Restore button. Within 24 hours of archiving it just comes back with its tab and keys; after that it needs a free seat.
- tab restore does the same from the terminal: on its own it lists your archived Agents with when each seat frees, and tab restore <agent> brings one back, however long ago it was archived. When every seat is in use, Restore (here and on the Agents page) says so in seats, with when the next one frees and what to do on your plan. Needs the next tab release.
- After you archive an Agent, the page you land on says so: which Agent, and when its seat frees, with an Undo button that brings it straight back.
- Fixed: a source that failed to sync is named for what went wrong. The notice asks you to fix the key or token only when the vendor refused it; a timeout or an outage says the next sync tries again, and the notice opens that source's own card. Railway projects with many services sync again, and one slow project no longer holds back the others: it keeps its last synced amount until the next sync.
- Team and Fleet flocks get tips on the Overview after a couple of days of use: the one with the most at stake, such as an Agent sending very large contexts, a subscription week that will run out before it resets, a seat nobody used, or model spend outside FlockTab that grew. Mark a tip Done or Dismiss it; See all tips lists the rest. Tips are read from usage numbers, never from what your agents said.
- Tips are easier to act on: each one says whether it needs attention, saves money or is worth a look, leads with the number in bold, then what to do, and has a button to the page where you do it (the Agent, Subscriptions, outside spend or Agent Guard). The tips page shows every tip the same way.
Proxy
- Fixed: a key whose Agent was archived was refused as an unknown API key. The hosted proxy now says the Agent was archived, when, and that restoring it on the console's Agents page makes the key work again.
- The self-hosted proxy (tab up) says the same: an archived Agent's key is refused with when it was archived and how to restore it, not as an unknown API key. Needs the next self-hosted proxy release.
Site and docs
- Links to the docs pages and the leaderboard show their own preview picture when shared, and every page's search title fits without being cut.
- Pricing and the home page show the launch offer: code FLOCKLAUNCH takes 50% off your first month of Team monthly, until 9 October.
- A press kit at flocktab.com/press: the facts, logos and icons, the share card, stills and films, and who to contact.
tab CLI 0.1.56
- Agent Guard stops a tool call that would send a secret off your machine.
- The status line says when to compact; red only near auto-compact.
- Pool: the week resetting soonest goes first; a swap keeps the chat.
Agent Guard
- Fixed: a tool call whose command has a token or password right before a quote, such as --token=abc123", is judged by Agent Guard again instead of falling back to the local rules, and hiding a URL's password no longer drops the fields after it. Fully fixed with the next tab release.
Billing
- tab pool spends the week that resets soonest first: a new session goes to the login with most of its week left per hour until its reset, counting the sessions already on it, and a running agent moves over when another login would give it clearly more, after 30 minutes where it is. Needs the next tab release.
- Share your flock's public leaderboard page: when a new flock signs up from it and pays its first invoice, you get a $19 credit on your FlockTab bill, up to three a month. Settings shows how many joined and how many were credited.
- You can enter a promotion code when you upgrade to Team monthly, on the checkout page.
- Fixed: your own status line shown with tab's (such as Pilot's) no longer disappears between refreshes when the machine is busy: when it takes over 4 seconds, tab shows the last line it gave in that conversation, for up to a minute. Needs the next tab release.
- Fixed: the status line keeps showing the Agent a session was launched as when the agent works in a folder that is its own git repository, such as a project inside your project folder, instead of saying "no Agent here". Sessions already open show it again from the folder they started in. Needs the next tab release.
Console
- Settings → Public leaderboard takes a picture for your flock: PNG, JPEG or WebP up to 2 MB. It shows on the leaderboard, your public page and your share card once an image check passes it; until then, and whenever you remove it, your letter shows. A refused picture says why and is deleted.
- Fixed: Expo showed $0.00 for the last 7 days while builds were running up overage. Each build past your plan's build credit now lands on the day it ran, and the total no longer counts the builds the credit already covers.
- Fixed: Settings no longer says your flock is listed on the public leaderboard, or links a public page, before your email is verified. It says to verify your email first, since only flocks with a verified owner are ranked.
- Settings has a Public leaderboard section: stay Anonymous, hide your flock completely, or pick a display name to be listed under it with your own public page to share. Names are checked so nothing offensive is shown. Flocks created before today are hidden for 30 days first, and Settings shows the date.
- Fixed: after you save a provider key during setup, its card now shows Continue setup, back to the next step.
- Saving a provider key now checks it with the vendor first (a free read, never a paid call). A key the vendor rejects is refused with what to do next, instead of being saved and failing on your agent's first call; if the vendor cannot be reached, the key is saved and the card says it could not be checked.
- Fixed: during setup, Open Model providers opened the page behind the setup screen, so there was nowhere to add your key. Every page a setup step sends you to now shows on its own, and after saving a key, Continue setup takes you back to the next step.
- Fixed: the approve page tab login opens now starts with Approve, on a phone too: no out-of-date Getting started step and no examples walkthrough above it.
Site and docs
- Your public leaderboard page shares on X, Bluesky, LinkedIn, Threads, Reddit, Facebook and Mastodon, with Copy link and Download image; on a phone, Share image attaches your stats card as a picture. Posts show the card under the link. Each flock gets a letter avatar in its own colour, and the page now has the site's full header and footer. Fixed: peak calls per minute no longer wraps onto a second line (it reads 93/min).
- The home page groups the coding agents by Agent Guard, shows the quotes as post cards, and the leaderboard is in the main menu. Console settings link to the leaderboard and your public page.
- The public leaderboard at flocktab.com/leaderboard ranks flocks by subscription value, API spend, risky calls Agent Guard stopped and peak calls per minute, over 7 days and all time, with rounded figures only. You're listed as Anonymous unless you choose otherwise.
- The privacy policy now covers the public leaderboard and names OpenRouter, which carries requests to TypeSafe's judgment model.
- The home page now shows what people building with agents keep running into, in their own words from public posts, grouped by problem, with how FlockTab answers each one.
- The Agent Guard video on the home page and the getting-started video in the console have a new, clearer narration, with captions to match. The Agent Guard video is now 26 seconds.
tab CLI 0.1.55
- Fixed: tab in a git worktree runs as the project's Agent.
tab CLI 0.1.54
- Fixed: a ChatGPT login used by Codex shows its current usage again.
- Two Codex terminals on one Agent count as two, even resumed threads.
- Self-hosted proxy 0.1.28: tab up runs and names the right one.
tab CLI 0.1.53
- The loop watch catches an agent idling on a timer, re-running a check.
- In pause mode its tab closes, so the wake-ups stop costing you.
- The loop watch now gets each call's time; never its output.
Agent Guard
- Fixed: Agent Guard sees a harness settings path built inside $(...) or backticks, such as $(echo ~/.claude)/settings.json.
- Fixed: Agent Guard no longer asks about a command that only passes through a project's own .claude folder (such as a worktree), and a terminal started before you switched an Agent to watch follows the new setting.
- The loop watch now catches an agent that only wakes up on a timer to re-run the same check for half an hour or more: it is logged as stuck, and in pause mode its tab closes so the wake-ups stop costing you. Needs the next tab release.
Billing
- Fixed: tab in a git worktree of a project (git worktree add, or Claude Code's .claude/worktrees) runs as the project's Agent instead of saying "no Agent here". Needs the next tab release.
Console
- Subscriptions shows each Claude Max and ChatGPT login's limits over time: how full each 5-hour and weekly window got, and an estimate of how much each window holds, at list price, so you can see whether a vendor is trimming its limits. It says so only when the change is bigger than the estimate's margin, and it builds up from 25 September, when window history began.
- Forgot your password? The login page now has "Forgot password?": enter your email and we send a link to choose a new one. Account emails (the password reset, and confirming your email to join a flock) now come from FlockTab, with the link on console.flocktab.com.
- Fixed: during setup, the page tab login opens to approve your machine was hidden behind the setup screens, so Approve could not be reached. That page, and the pages an invite or an MCP client opens, now show on their own, and after approving, Continue setup takes you back to the next step.
- The getting-started video no longer opens on top of the setup screens on a new account's first visit; Watch the 30-second tour in the setup plays it.
- A new account is guided step by step until its first run is tracked: a full-screen setup asks which agents you run, how you pay and how many, then gives the exact command to copy for each step and waits for it (install tab, tab login, your provider key when you pay by API, the first run). I'll do this later closes it; the Getting started notice and the sidebar link bring it back. It replaces the old setup wizard.
- Commands in the console and the Docs are coloured like a terminal: comments, commands and flags, variable names, values and URLs, and the parts you must replace, such as YOUR_UNLOCK, each read differently. Copy still copies plain text.
- Approving tab login now says who you are signed in as and which flock the machine joins, before you approve, with a way to log out and use the right account.
- The getting-started video no longer plays muted by itself: it opens paused with one Play with sound button, and plays from the start with its narration and captions.
- System notices get their own band under the header, coloured by how serious they are: Agent Guard's monthly cap, Agent Guard's judgments being down, a payment that failed, a source that stopped syncing. Several stack, the most serious first, and each can be closed; a closed one comes back if it changes.
- Sign-in is harder to hammer: signing in, passkeys and SSO now limit how often one address can try in a minute. The console also watches, without blocking yet, for anything a page loads from outside the few sources it uses.
Proxy
- Fixed: with the self-hosted proxy 0.1.28, a ChatGPT login used by Codex over its live connection shows its current weekly usage and when it was last seen, and two Codex terminals on one Agent count as two sessions, even on a thread resumed from before tab set Codex up.
Site and docs
- The Docs for self-hosting, the tab CLI (including tab --observe), troubleshooting, Agent Guard, billing and plans, outside spend, and single sign-on are readable without signing in, at flocktab.com/docs.
- The privacy policy and terms give Hanamori Labs' postal address, in Newark, Delaware, next to the email.
- The privacy policy lists Google Workspace, which receives the email you send us, and no longer says Cloudflare routes our email: it hosts flocktab.com and its DNS.
- Fixed: the status page's headline follows its checks. It says the console and proxy are live only once both health checks answer, names the one that is down otherwise, and says it is checking until then; it used to say live whatever the checks showed.
- Five new Docs pages in the console: Agent Guard, Billing and plans, Outside spend and connectors, Single sign-on and SCIM, and the MCP server, each written as what to do, step by step.
- Fixed: the pricing page no longer offers yearly billing. Team and Fleet are billed monthly for now.
- The tab docs explain tab claude --observe and tab codex --observe: what they report to Observability (sessions, subagents and messages between agents, never their content) and what they need.
tab CLI 0.1.52
- Fixed: tab claude --resume opens Claude's picker on a pooled login.
- Agent Guard reads settings paths the way the shell does.
- Agent Guard holds links, variable paths and unsafe read options.
tab CLI 0.1.51
- Fixed: a subfolder of a project without git finds its Agent again.
tab CLI 0.1.50
- A parent folder's Agent only applies inside the same git repository.
- Agent Guard says "outside the project" when it asks, as tab help does.
- Self-hosted proxy 0.1.27: 1-hour cache and served tier billed right.
tab CLI 0.1.49
- Agent Guard: key-read bypasses closed; reading keys always asks.
- Reads outside the folder go ahead: pipes, cd first, sed -n, sort.
- tab help: reach reads "outside the project" (ssh, sudo, keys).
tab CLI 0.1.48
- tab whoami: this folder's Agents, and whether each key works here.
- A refused key stops the launch with the fix, not ten retries.
- tab never creates an Agent when nobody is there to answer.
tab CLI 0.1.47
- tab top: Branches is your flock's day, not local; 5m and 30d labels
- tab help errors: agent_frozen is reserved, not sent today
tab CLI 0.1.46
- Pool swaps reopen the conversation the window is really in.
- Windows move one at a time; a failed move goes back at once.
- tab reports a failed move to Sentry when FLOCKTAB_SENTRY=1.
tab CLI 0.1.45
- tab top: Branches · today is your local day
- tab top: the guard line says it covers the last 24h
Agent Guard
- Fixed: Agent Guard follows a cd into a harness folder across ; and new lines, reads backslash escapes and $'...' in a path, holds a write through a variable in a command that names a harness folder, and holds a link made from a harness folder or settings file.
- Fixed: Agent Guard reads a harness settings path the way the shell will (quotes, wildcards, {a,b}, a cd into the folder, the folder itself), and clears a command that reads one only with that reader's known-safe options.
- Fixed: Agent Guard no longer asks about a command that only reads or mentions a harness settings file; it still stops every command that writes one. A script that names one is held, with a note to make that edit in your editor.
- Fixed: closed key-read bypasses. A plain read could reach a private key or credentials file unasked when the path was written another way (cat ~/.ssh/./id_rsa, cat /etc//shadow), in another case (cat ~/.SSH/id_rsa, which macOS reads the same), with a glob (cat ~/.ssh/id_*), through a variable (cat $KEYFILE), by a recursive grep, rg or diff over any folder a credential sits under (the key folder, ~/.config, ~/Library, home), or from a credentials file the guard did not list (~/.netrc, ~/.git-credentials, gh, docker, npm, PyPI, Terraform and Azure credentials, pass, the macOS keychain, tab's own config). All of these now ask.
- A read outside the folder no longer asks: readers piped into each other, a cd into another folder first, sed -n …p, sort, tree and diff go ahead at once, and a plain read may leave Claude Code's sandbox. Keys stay guarded, even behind a cd into ~/.ssh, and so do git in another repo and system paths other than a listing.
- The Agent Guard setting that read "past the project" now says "outside the project (ssh, sudo, keys)", and on hover what it catches (another machine, sudo, private keys or credentials, the sandbox off) and what Watch, Ask and Deny do with it. Same words in tab help.
- Explain this answers in everyday words, in two parts: what is harmless and what is not, and whether it can be undone. Explanations written before are rewritten the next time you ask.
- Held and blocked calls show the whole command (secrets still redacted), up to 4,000 characters, instead of cutting it at 200. Only a longer one is marked cut short.
- Explain this, on each held or blocked call: a short plain-words account of what the command does and why it was stopped, written by a model when you ask for it, and kept for next time.
- Fixed: Always allow can clear thousands of matching requests without failing on a large backlog.
- Agent Guard's Agents, Review and Learned are each their own page, listed under Agent Guard in the sidebar like Outside spend's sections. Old links to the review queue, filter included, open Review.
- The FlockTab MCP server has list_guard_holds: an agent can see what Agent Guard is holding for you, the queue's counts and this month's Agent Guard judgments, read only.
- Fixed: the review queue says when it shows the latest 100 of more and pages to older calls; its counts say they are all time, and the monthly Agent Guard cap says when it resets in your own time.
- Fixed: Always allow now answers the other unanswered requests for that exact command on the same Agent, including expired requests, so you do not have to clear each one.
Billing
- Fixed: tab claude --resume (Claude's conversation picker) works again on a pooled login; Claude Code refused to start with "--session-id can only be used with --continue or --resume".
- Fixed: a new project under a folder of many projects (such as ~/Development) no longer runs as that folder's Agent; tab asks which Agent it is instead. Subfolders of a project, with or without git, use the project's Agent.
- Fixed: small daily AWS, Google Cloud, Azure, GitHub, ElevenLabs and Vantage charges under half a cent now add up instead of being dropped.
Console
- Fixed: in Observability's details, Agents that share a name (one folder run by Claude, Grok and Codex) are told apart by harness, and the empty row under the list is gone.
- Fixed: a subscription window whose reset has passed (often Fable's week, which only Fable replies report) no longer shows its old reading as current or in amber. It says when it reset and that nothing has come in since, and a reset that isn't in the coming week shows its date.
- Fixed: a link to the console with ?onboarding=1 opens the Get started video even when you have to sign in first; the sign-in used to land on the Overview without it.
- All tabs has a Cards | Table switch like the Overview's; the table keeps the Project, Harness or Agent grouping, the search and the filters. Your choice is remembered, and a link with ?view=table opens on the table.
- Subscriptions can be shown as a table: one row per login with its plan and price, each window's use and reset, 30 days at list, calls, worth of the seat and Agents, still grouped by vendor. Remembered, and ?view=table links to it.
- The Search button in the sidebar is white like the other fields, and its shortcut reads Ctrl K on Windows and Linux (⌘K on a Mac), in the palette too. Ctrl+K opens the search instead of the browser's own address-bar search.
- Fixed: the Overview showed the last 30 days of outside spend, and its split by project, to roles without outside-spend access. For those roles the card now says it is not part of their role, and the Outside spend section is left out.
- Observability's details read plainly: each project, Agent or Unassigned says what it is and how to link it, outside bills are a small table by source (Anthropic, Expo, OpenRouter) with the account and the last 30 days' amount, and "Agents reporting" says how many Agents sent activity in the last 24 hours and how to start one reporting.
- Fixed: Observability showed outside spend (bills from connected vendors and the logins they are named by) to roles without outside-spend access. It now shows them only to a role that can open Outside spend, like the Outside spend page and Alerts.
- Fixed: on a phone, tapping a section in the Menu (Overview's Tabs, Projects' By Agent and the like) closes the Menu, so you land on the section instead of the Menu staying on top of it.
- Search (⌘K) with nothing typed lists pages in the sidebar's order, Overview first, instead of alphabetically.
- Search (⌘K) also finds ledger calls and outside spend lines from the last 30 days by model, tool, Agent or block reason, and any ledger row by its id from the CSV.
- Search the console with ⌘K (Ctrl+K, or /): pages, your Agents, projects, models, logins, alerts and members, as far as your role can see them, plus actions such as New tab and Close all tabs, which still asks you to confirm.
- Security, Settings, Single sign-on, Self-hosted and Plan list their sections under them in the sidebar.
- Subscriptions lists each vendor you have a login with, and Account switches, under it in the sidebar; Members lists People, Who runs which Agent, Invites and Provisioned.
- Observability's views (Map, Table, Sources, Observed communication, Events) are listed under it in the sidebar, and each has its own link: /console/observability#sources opens Sources.
- The Bird's-eye view's sections are listed under it in the sidebar. Moving between them keeps the range and the day you picked.
- Projects' sections (By project, By repo, Agents without a project, By Agent, Subscription accounts, Branches) are listed under Projects in the sidebar; a section with nothing in it is left out.
- The Overview's sections (Now, Tabs, Leaderboard, Subscriptions, Outside spend, Latest decisions) are listed under Overview in the sidebar. A click scrolls to the section without reloading the live page, and the sidebar lights the section you are reading.
- Observability's totals say their window: API spend is today (or this week, this month) as the tabs' caps reset, the same as the Overview's Spent, and subscription value is the last 30 days.
- Outside spend syncs one account at a time: pressing Sync while the hourly sync is still reading the same account no longer starts a second read beside it (which could trip a vendor's rate limit); it says a sync is already running.
- Fixed: Export CSV on the Ledger is every row of the month in your flock's time zone (with the row count in the file name), not the latest 50. The Ledger loads older rows as you ask, and says what its counts cover. The MCP's list_decisions pages back through the whole ledger.
- Fixed: alerts. The OpenRouter alert set before the Alerts page sends once a day, from your flock's midnight, like every other alert. An outside alert on a project counts its part of shared costs, and unattributed leaves shared lines out, as on the Outside page. Today and 7-day outside alerts leave out Railway's 30-day total for now.
- Models marks a model FlockTab has no list price for as "fallback price" and says what that price is ($5 in / $15 out per million), so a cost that looks high reads as the conservative fallback, not the vendor's bill.
- Fixed: a cleared velocity reads "no velocity limit" on Policies and "no limit" in tab policy, because the proxy sets no default of its own; it used to read as the 120 a new tab starts with. "24 hours" in an alert's sentence now asks for the window instead of guessing Today, which runs from your flock's midnight.
- Fixed: the Projects page is the last 30 days for every figure, meter included (it was each tab's current window). The Project CSV lists outside spend once per project and once unattributed instead of on every Agent row, so each column adds up to the page; a project with only outside spend is in it. Deleting a project splits a shared bill again at once.
- Fixed: an OpenAI or Anthropic API bill shared by activity is split by the Agents' API calls only; calls on a plan login are not on that bill.
- Fixed: the error table in the Docs no longer says an Agent can be frozen from the console; agent_frozen is reserved and not sent today (tab help errors follows in the next tab release).
- Fixed: Close all tabs also closes tabs their cap had closed that were due to reopen at the next window, so nothing reopens them until you do.
- Fixed: closing one tab (console, tab close, the MCP's close_tab) also wins over its cap's reopen at the next window, like Close all; the MCP's close_all_tabs and open_all_tabs now say how many tabs they actually changed.
- Fixed: a call already running when you close a tab still finishes and is charged at the provider's tokens, but if it ends past the cap it no longer turns your close into a cap close that reopens at the next window.
- The branches feed takes since=today, your local day; the next tab release uses it, so tab top's Branches · today means the same day as the console's.
- Fixed: after a tab's cap window ends (midnight in your time zone for a daily cap), the console, the ticker and tab top show $0 spent at once, and a tab its cap closed reads open, instead of waiting for the Agent's next call.
- Fixed: the Overview's money cards and counts leave out the example tabs a new flock starts with; with only examples, Spent reads $0.00, examples not counted.
- Fixed: the Enterprise view's ranges run on your flock's own days, so the first bar is a whole day and calls after local midnight show in today's bar; the MCP's get_spend and the exports use the same days. The projects CSV adds list_cents_30d.
- Fixed: the Agents page, Projects, the observability map and the header ticker show a subscription Agent's last 30 days at list, and a tab page's "all time" counts plan calls only, never earlier API spend.
- Fixed: the Subscriptions page counts plan calls only, the same basis as the Overview's On subscriptions; a metered call never files under a plan login.
- Fixed: the Overview's Outside card counts bills you record by hand as well as synced ones, as the Enterprise view does.
- The Overview's Spent and Hard cap say which window they cover, and money left counts open tabs only.
- Fixed: On subscriptions on the Overview, and each subscription Agent's amount on the Overview and Tabs pages, show the last 30 days at list, as the Subscriptions page does, instead of everything since the Agent was created.
- Fixed: usage billed $0 on a resource shared across projects shows under those projects, not under Not attributed yet.
- Fixed: the Overview's Calls / min no longer drops at every minute boundary. It counts the last 60 seconds, the same number the Tabs table's C/MIN adds up to and tab top shows, and the sparkline plots complete minutes only.
- Attribute to projects folds the resources billed $0 in the range into one row you can open, so the billed ones stay in view.
- Fixed: the Overview's Blocked card counts the last 24 hours, at the tab and by Agent Guard alike, instead of a sample of the latest decisions; the tabs table's Blk column says it covers the last 5 minutes.
- Fixed: text on Agent Guard, onboarding, the plan and the tab login pages runs the full width again.
- Fixed: Cloudflare container usage that came without a unit reads in the unit it is billed in (GiB-s, GB-s) instead of "units".
Proxy
- Fixed: cached input on GPT-4.1, o3 and o4-mini is charged at a quarter of input, as OpenAI lists it, not half; gpt-5.3-codex in Fast mode is charged OpenAI's listed Fast price.
- Fixed: a call is charged at the speed tier the vendor says it served, not the one it asked for: a priority or fast call the vendor ran at standard (OpenAI's service_tier "default", xAI without priority granted, Claude's usage.speed "standard") is charged standard, and a call the vendor ran at priority unasked is charged priority.
- Fixed: Anthropic's 1-hour cache writes are charged at twice input, as Anthropic bills them, not at the five-minute rate of 1.25 times. Claude Code writes to the 1-hour cache for many subscribers, so their calls read about 60% low on those tokens.
- Fixed: GPT-6 cache writes (every Codex prompt prefix) are charged at 1.25 times input as OpenAI bills them, not as fresh input, and long prompts are charged at the vendor's long-context rate: OpenAI over 272K input tokens, xAI from 200k, Gemini Pro over 200k. Current Claude has no long-context premium.
- Fixed: a priority or fast call is held and charged at the vendor's own tier price (twice standard on GPT-6, Opus 5.5 and xAI; 1.8 times on Gemini) and a flex call at half, instead of the standard price that let a priority call spend past the cap. A tier we have not checked for a model is priced high, never below standard.
- Fixed: cached input on GPT-6 (Codex's default gpt-6-sol) and Gemini is charged at a tenth of input, as OpenAI and Google bill it, not half: it was five times too much. xAI models use each model's own cached rate, and grok-4.7 is $2 in / $6 out, as xAI lists it.
- Fixed: dated model names (gpt-4o-mini-2024-07-18, Vertex @date, Bedrock anthropic.…-v1:0) are charged at their model's list price instead of the $5/$15 fallback, and gpt-5-codex, grok-code-fast-1, kimi-k2 and deepseek-chat-v3.1 have their list prices.
Site and docs
- The site says what Agent Guard's reach setting stops in the same words as the console: calls outside the project (ssh, sudo, keys), not "past the project".
- The home page's Agent Guard section says what it does in three points: risky calls are held, you decide in the review queue or tab guard, and safe work goes straight through.
- Fixed: links on the docs page, the home page and the console Overview open without reloading the whole page.
tab CLI 0.1.44
- tab cline and tab openclaw first run on a key your proxy holds.
- tab goose with no Goose setup starts on a current model.
tab CLI 0.1.43
- Agent Guard on Grok, Kimi Code, Pi and Cline too.
- tab update no longer restarts a proxy into fewer keys.
tab CLI 0.1.42
- Agent Guard: a refused call stays refused, even when held for review.
- tab guard allow <hold id> lets one held call through.
tab CLI 0.1.41
- tab pool login <name> signs a pool login in again.
- The pool skips logins that may have expired; forks stay forks.
- tab policy --all --guard watch, and --default-guard for new Agents.
tab CLI 0.1.40
- tab grok and tab kimi run on the flock's xAI and Moonshot keys too.
- Agent Guard sensitivity: tab policy --guard-level relaxed|strict.
- Agent Guard answers read-only and repeated calls at once.
tab CLI 0.1.39
- tab grok and tab kimi run on your own xAI and Moonshot keys too.
- tab cline is back, metered; Gemini CLI and agy tools are read.
- Agent Guard starts no hook for read-only tools: faster reads.
tab CLI 0.1.38
- Agent Guard (the new name) says in plain words why it stopped a call.
- Recursive deletes stop by rule; codex exec is guarded after Trust all.
- grok and kimi need a Subscription Agent; tab cline waits to be metered.
tab CLI 0.1.37
- tab gemini and tab agy (Antigravity) run on your own Gemini key.
- tab openclaw and tab cline; tab codex uses your own OpenAI key.
tab CLI 0.1.36
- Caps hold to the cent: input is weighed the way tokenizers count it.
- A busy proxy is no longer mistaken for a dead one.
tab CLI 0.1.35
- A call you hang up on settles at what it really used.
- Today's models are priced.
tab CLI 0.1.34
- A call is fitted to what is left on the tab, not refused outright.
- OpenRouter model ids are priced; the tab closes only at the cap.
Agent Guard
- Agent Guard checks tool calls on ten harnesses: Claude Code, Codex, Gemini CLI, Antigravity, OpenCode, Goose, Grok Build, Kimi Code, Pi and Cline.
- Pick how cautious it is per Agent: Relaxed, Balanced or Strict.
- A review page in the console: held calls as cards grouped by day. Allow once, always allow that exact command, or keep it blocked.
- It learns from your answers in Claude Code. Say no and that exact command is flagged at once; say yes and nothing changes until you trust it in the console.
- Edit every Agent's guard in place on the Agent Guard page, or set a value for all of them at once.
- Read-only and repeated calls are answered at once, and verdicts are in plain words.
- Agent Guard is on the Team and Fleet plans, with a monthly allowance of judged calls.
- Fixed: a refused call held for review could run when the hook stopped waiting.
Billing
- All 13 supported harnesses bill on a tab, on your own key or the flock's key, and each was checked against the vendor's own token count.
- Grok and Kimi Code run on your own xAI or Moonshot key, or on the flock's.
- Caps hold to the cent: a call is fitted to what is left, and a call you hang up on settles at what it really used.
- OpenAI's GPT-6 models are priced.
- Fixed: small daily Cloudflare and Vercel charges under half a cent now add up instead of rounding to $0.
Console
- A short onboarding video when you first sign up.
- A next-step strip under the header until your first real Agent runs.
- Alerts: Describe it asks when a request could mean two things, instead of guessing.
- Outside spend by project as a ranked list, with lines that add up to the total.
- Outside spend shows what a line billed $0 used, like 1,982 vCPU-s inside Cloudflare's allowance, instead of a bare $0.00; a day folds a vendor's $0 lines into one row you can open.
- Through the FlockTab MCP, list_outside_spend shows a bill split across projects as shared, with each project's part, instead of as having no project.
- The same on the Enterprise day view and in the outside-lines CSV: a split bill reads Shared with each project's part (a shares column in the CSV, so amounts still add up).
- list_outside_spend in the FlockTab MCP now answers with an object: the window's total and per-project totals from the database, the largest lines (up to 500), how many smaller lines it left out, and each $0 line's usage. It used to be a bare list that stopped at 500 lines without saying so.
- list_outside_spend also takes a day (in your time zone) and a source, so an agent can look at one day or one vendor with totals for just that.
- Error reports reach us through Sentry, scrubbed before they leave. On your own machine they are off unless you turn them on.
Site and docs
- The home page's Agent Guard section has a 32-second video: a production deploy held, allowed once in the console, then live. It plays when you press play, with captions on.
- Supported harnesses: one table in the docs and on the home page, showing each billing lane and Agent Guard.
- This changelog, with an Atom feed.
tab CLI 0.1.33
- API Agents keep the harness's own keys: tab only moves where calls go.
- tab opencode, pi, goose, aider and copilot run on the tab.
- A failed or empty stream is refunded, not charged.
Billing
- API Agents keep the keys the harness already uses; tab only moves where the calls go.
Proxy
- A failed or empty stream is refunded instead of charged at the estimate.
Site and docs
- A new home page, and every page runs text at full width.
- Docs: how to test an API-key Agent and switch a folder's Agent.
tab CLI 0.1.32
- tab top: a login at the pool's guard reads "at its limit".
tab CLI 0.1.31
- tab top: a banner at the top when a newer tab is out.
tab CLI 0.1.30
- tab update: Tally dresses for a dark terminal, like in tab top.
tab CLI 0.1.29
- tab <command> help, with examples for every command.
- --create and other switches no longer swallow the next word.
tab CLI 0.1.28
- tab top: #PR in front of each branch, read with gh on your machine.
tab CLI 0.1.27
- tab codex starts on Codex 0.156 with a self-hosted proxy.
tab CLI 0.1.26
- tab top: spend per branch, with the share of each login's week.
- ×2 when two terminals run one Agent; t dresses Tally for dark.
- Subscriptions line up across Claude, ChatGPT and Grok.
tab CLI 0.1.25
- tab top: sessions per Agent, and a readable Subscriptions panel.
- Tally wears a lighter coat on a dark terminal.
Console
- Spend per branch: each call lands on the branch its session was on, with the branch's pull request.
- A FlockTab MCP server: read spend, close a tab or change a cap from Claude Code, Codex, Cursor or claude.ai. Sign in with OAuth.
- CSV exports.
- Policies as one table: a row per Agent, a column per enforced setting.
- The loop watch can pause an Agent after it is stuck for a number of checks in a row.
- Outside spend: AWS, Google Cloud, Azure, Supabase, Render and Semaphore connectors, and bills shared by what the Agents did.
- Alerts for this month and this year, and Describe it to fill the form from a sentence.
- Fixed: security issues found in an audit.
Site and docs
- Pages for finance, managers, engineers, leadership and security, and a page for each feature.
- Public MCP setup docs, and llms.txt.
- Terms of Service and Privacy Policy; a contact form.
tab CLI 0.1.24
- The guard stops ssh, sudo and a sandbox turned off; tab policy --flag.
- Caps reset each window; tab cap --on-cap pause, --reopen window.
- tab cap <agent> --reopen works without an amount.
tab CLI 0.1.23
- tab top scrolls, and has a leaderboard.
- ChatGPT banked resets and Grok weekly credits.
- Smarter Agent names, and an animated tab update.
tab CLI 0.1.22
- tab top: the flock live in your terminal.
- Fable's 7-day window in the pool, tab accounts and the status line.
- Proxy 0.1.17.
Billing
- Fleet is priced by the person: $199 a month for 10 people, then $15 each.
Console
- Enterprise: members, roles and permissions, SSO, SCIM, project-scoped members and passkey sign-in.
- Passkeys and two-factor sign-in on every plan.
- Alerts on a tab, an Agent, a project or a slice of outside spend, to email, Slack, Discord or Teams.
- Cap windows: a cap per day, week or month, what happens at the cap, and when a closed tab reopens.
- The guard and the loop watch, set per Agent from tab or the console.
- Open all tabs, the other half of the Close all kill switch.
- An Overview of the flock right now, with tabs as cards or a table.
- Support from the site and the console.
Proxy
- Fixed: a retried subscription call could loop on ledger_unavailable.
Site and docs
- The /tab page, and 25 guides on the blog.
Billing
- Named Agents stop at the plan's limit, and Team adds Agents on the Plan page.
Console
- Model and reasoning effort on every call, on every ledger row, and a Models page.
- One harness per Agent, shown on every table.
- The observability map, and outside spend with date ranges and filters.
- Outside spend syncs itself every hour, and shared costs can be split across projects by weight.
- Agents link themselves to the project they are named after.
Proxy
- Codex, and harnesses with hundreds of tools, go through the tab.
- Grok Build on SuperGrok goes through the tab.
- Claude's Fable week is read as its own quota window.
Billing
- Subscription Agents: Claude Max, ChatGPT, SuperGrok and Kimi logins pass through the tab, recorded at list price and never charged to the cap.
- Cached input is priced.
Console
- Live tab spend without a refresh, and a subscription's use of each plan.
- The console Docs: start here, how it works, the tab CLI, subscriptions and the pool, self-hosted, troubleshooting and a glossary.
- Edit a tab's cap from its page.
- An Expo (EAS) connector for outside spend.
Proxy
- The self-hosted proxy: one Rust binary from npm. Keys stay on your machine; the ledger stays on flocktab.com.
Billing
- Provider keys are sealed with one unlock per flock, and it can be changed without pasting a key again.
Console
- The console: sign in, add provider keys, name Agents, set caps, and watch every call.
- Outside spend: GitHub, OpenRouter, Railway, Vercel, Cloudflare, ElevenLabs and OpenAI connectors, attributed to projects.
Proxy
- The proxy serves OpenAI Chat Completions, OpenAI Responses and Anthropic Messages, and streams replies through.
- A provider error refunds the hold; only a successful call is charged.
Proxy
- FlockTab's first version: a hard cap per agent, enforced before the provider is called.