For engineers
One command in front of your agents. Nothing else changes.
tab claude starts Claude Code through FlockTab with a hard cap on its tab. Same harness, same prompts, same login or key. If it loops at 3am, it stops at the cap instead of at your card limit.
Jev's guard in the agent's own session: ordinary work runs, irreversible calls are stopped before they run.
The failure modes you already know
- An agent retries a failing tool call all night, and every retry is a paid request.
- A helpful agent force-pushes, drops a table or deploys to production because nothing asked first.
- Budget tooling wants an SDK in your code, or a gateway that sees all your prompts.
With FlockTab
Drop-in, no SDK
npm i -g @hanamorilabs/tab, tab login once, then tab claude, tab codex, tab grok or tab kimi. Any OpenAI- or Anthropic-compatible client works with a base URL.
Your keys or your subscriptions
API agents use your provider key. Subscription agents use your own Claude Max or ChatGPT login, gated and recorded, never charged. Several logins pool by headroom.
Self-host in one command
tab up runs the proxy on your machine: prompts, replies and keys go straight to the provider. Only model, tokens and cents reach flocktab.com.
Refusals your agent can read
A refused call is an HTTP error with a reason, 402 tab_closed or 429 velocity, before the provider is ever called.
- Streaming passed through untouched
- Self-hosted proxy is one static binary
- MCP server to read and close tabs from your agent
- Free for three agents
Put a tab on your agents tonight.
Two minutes with the tab CLI. Set a cap, run the agent, sleep.