
Squidbrake
Change control for AI agents: sign-off, undo and a record
60 followers
Change control for AI agents: sign-off, undo and a record
60 followers
Every action your AI agents take (commands, refunds, emails, database changes) goes on a tamper-evident record. Routine ones run; risky ones wait for a second person, who sees what led to it, with an undo. Claude Code, Cursor, Codex, MCP. Free, open source.
This is the 2nd launch from Squidbrake. View more
Squidbrake
Launching today
Your agents can do more than suggest. Squidbrake lets them ship code, change databases and handle refunds, with a person signing off only where it matters. Routine work runs, risky work waits for approval, catastrophic work is blocked, and every decision lands on a tamper-evident record your auditors can read. Works across Claude Code, Cursor, Codex and MCP.







Free Options
Launch Team

AWS Startup AdvisorPersonalized AWS guidance that meets you where you’re already building
Promoted

Hi Product Hunt 👋
We're Karishma and Pulkit, the founders of Squidbrake.
Talking to teams building with agents, we kept hearing the same thing: the agent works well in the demo, then gets held back from production. It's not that it can't do the work. It's that nobody has decided which actions need a person, and nothing records who approved them. So agents are stuck suggesting when they could be doing.
Squidbrake fixes that. Routine actions run on their own. Risky ones, like a prod deploy, a DROP TABLE or a refund, wait for a person to approve. Catastrophic ones are blocked. Each action is judged by the steps that led to it, not just the action itself, and no AI model makes the call. Every decision goes on a tamper-evident record.
By the numbers:
10,000+ downloads on PyPI in the six weeks since 1 October
68 of 71 harmful steps stopped in full agent sessions, while only 3 of 41 routine steps were held for a person
42 of 42 reported bypasses stopped. Every way around Squidbrake that someone sends us becomes a test that runs on every release
17 of 20 harmful actions stopped when we replayed 16 real public agent incidents (the 3 misses are written up)
What's new since our last launch:
Hosted version: sign up and get your own dashboard, with no server to run
Observe mode: see what Squidbrake would hold before you turn it on
Starter rules for Gmail, Jira/Confluence and Kubernetes, plus approval alerts in Telegram
Many more risky commands caught, with four new incident replays
Examples for Pydantic AI and the Vercel AI SDK
It works across Claude Code, Cursor, Codex, Gemini CLI and any MCP tool.
One question for you: which action would you never let your agent run without a person signing off? We'd love to hear yours, and what's stopping your agents from going to production.
Squidbrake lets AI agents run on their own in prod, while anything risky (deleting data, force-pushing, moving money) waits for a human to approve. Works with Claude Code, Cursor, Codex, MCP tools and more.