Admins enable each coding agent.
Claude Code, Cursor, and Codex are switched on per org — an agent nobody enabled cannot open a session.
Claude Code, Cursor, and Codex work with real repos and systems — give them brokered, revocable access instead of raw keys.
Coding agents read repos, call APIs, and ship changes — powered by whatever credentials live on the laptop.
Open a pull request for the retry fix.
github.open_pr · acme/checkout #481Opened through Alex’s own GitHub connection. The laptop never held a provider key — the credential was attached server-side for this one call.
Pull in the payments service for context.
github.read · acme/payments-coreAlex has no access to that repository, so the agent working for him has none either. Refused at the broker; GitHub was never called.
Ship it to production.
deploy.release · productionA named human approver has to say yes first. The session stays open and nothing ships while the request waits.
Fine — I’ll take the deploy myself.
Illustrative session · the developer sees every verdict as it happens, and each one lands on the same audit trail.
$ coding-agent --connect fabriqopening sign-in…signed in as alex@acmesession open · tools from your own connectionsNo provider key is stored here. Credentials stay in the vault and are injected server-side at call time.
alex@acmeclaude-codeopenpriya@acmecursoropensam@acmecodexclosedIllustrative sessions. Every broker session is attributed to the developer who started it.
Claude Code, Cursor, and Codex are switched on per org — an agent nobody enabled cannot open a session.
Copy-paste config; provider credentials stay in the vault and are injected server-side at call time.
Every active broker session shows in the console, attributed to the developer who started it.
Switch the coding agent off and no laptop can establish a session again — nothing to rotate, nothing to chase down.
The agent works through that person’s connections, so it can’t reach a system they haven’t connected themselves.
Developer, agent, tool, verdict, timestamp — the same ledger as the rest of the fleet.
Explore the context attached to every request.
coding-agent has a defined purpose and a bounded set of tools.
The request acts for developer, using that person’s permissions.
The policy applies to the requested action within GitHub, before the tool executes.
Claude Code, Cursor, and Codex through the local broker — plus anything MCP-compatible on the same governed endpoint.
Switch the agent off for the organization. From that moment no laptop can establish a session, and because the broker re-authenticates rather than holding a standing key, its automatic reconnect is refused too. Revoking a developer’s live session from the console closes out what is already open.
No — setup is a copy-paste config and one sign-in. After that the agent works exactly as before, just through a governed endpoint instead of raw keys.
No. Sign-in is an OAuth login — a device code flow for headless machines — and provider credentials stay in the vault, injected server-side at call time. Nothing a developer holds locally can reach a provider on its own.
The tools the developer has connected, through the org’s hub. It inherits that person’s access rather than carrying standing credentials of its own — so offboarding the developer offboards their agent with them.
Yes — live broker sessions are visible in the console as they happen, each one attributed to the developer who started it.