# Squidbrake

> Squidbrake is change control for AI agents: agents propose, people approve. It checks every tool call an AI agent makes against your rules and what the agent already did, and allows it, holds it for a second person (with what it will change shown first, such as rows deleted or resources destroyed), or blocks it, before the call runs. Approved deletes can be undone, and decisions land in a tamper-evident log with a one-page evidence pack for auditors. Decisions come from rules, not an LLM.

Squidbrake runs where you run it: a laptop, Docker, or your own cloud. It is backed by Entrepreneur First.

## Start here

- [This site as markdown](https://squidbrake.com/index.md): what Squidbrake does, the incidents it was tested on, how policy works, measured latency, FAQ
- [Quickstart](https://squidbrake.com/docs): install, connect agents, write a policy, block a call
- [PyPI package](https://pypi.org/project/squidbrake/): `pipx install squidbrake`
- [Pricing](https://squidbrake.com/pricing): Free (one person, on your computer), Team $499/month with a 14-day trial, Business and Enterprise priced with you; what is not built yet
- [Security](https://squidbrake.com/security): where it runs, what it stores, how secrets and keys are kept, what it can't stop, compliance mapping (SOC 2 CC8.1, ISO 27001 A.8.32 and A.8.15, EU AI Act Art. 12 and 14), reporting a vulnerability
- [Changelog](https://squidbrake.com/changelog): every release, newest first

## What it connects to

Coding agents (Claude Code, Cursor, Codex, Gemini CLI, VS Code Copilot, Antigravity), agents that connect to MCP by URL (ChatGPT and claude.ai connectors, Devin, n8n), agent frameworks (OpenAI Agents SDK, LangChain, LangGraph, CrewAI, PydanticAI, smolagents, Google ADK), and any MCP server. Support, finance and ops agents go through the same rules as coding agents.

## Evidence

- [Benchmark write-up](https://squidbrake.com/benchmark): 12 public AI-agent incidents replayed against the shipped rules (last run 21 Oct 2026, v0.14.2). 14 of 15 harmful actions stopped, the one miss documented. Code: [incidents](https://github.com/batrapulkit/squidbrake/tree/main/incidents), `python incidents/replay.py`.
- [Session benchmark](https://github.com/batrapulkit/squidbrake/tree/main/bench/chains): 62 agent sessions, 113 steps. Squidbrake on whole sessions stopped 65 of 66 harmful steps and held 3 of 41 routine ones; an allowlist stopped 53 and held 10; asking for everything held 40.
- [Default policy](https://github.com/batrapulkit/squidbrake/blob/main/rules.yaml): the rules Squidbrake ships with

## Guides

- [How an AI agent deleted a production database, and how to stop it](https://squidbrake.com/blog/ai-agent-deleted-production-database)
- [Human-in-the-loop approval for AI agents](https://squidbrake.com/blog/human-in-the-loop-approval-for-ai-agents)
- [MCP security checklist](https://squidbrake.com/blog/mcp-security-checklist)
- [EU AI Act Article 14: human oversight for AI agents](https://squidbrake.com/blog/eu-ai-act-article-14-human-oversight-ai-agents)
- [Squidbrake vs Kontext](https://squidbrake.com/blog/squidbrake-vs-kontext): human approval, open source, MCP, self-hosting, price

## Contact

- Email: hello@squidbrake.com
- Book a 30-minute demo: https://calendly.com/pulkitbatra2024/new-meeting-1