Cirvix

Runtime authorization for AI agents. Default-deny policy for every tool call your coding agent makes. Open source (Apache-2.0).

[Cirvix @cirvixai]

Your AI coding agent can read .env, run shell commands and call every MCP tool you connected.

Nothing in the agent decides whether it should.

Cirvix does. Every tool call gets ALLOW, HOLD or DENY before it runs. Locally. Default-deny.

[Cirvix @cirvixai]

If your agent can drop your prod DB, one day it will. Not malice. Probability.

[Cirvix @cirvixai]

Claude Code, Cursor, Codex, Cline, Gemini CLI. One policy for all of them. Cirvix.

[Cirvix @cirvixai]

Quick test: can you list every command your AI agent ran yesterday? Most teams can't.

[Cirvix @cirvixai]

Agents don't need to get smarter. They need brakes. That's Cirvix.

[Cirvix @cirvixai]

Cirvix sits between your agent and every tool call. Allow, hold, or deny before it runs. Open source.

[Cirvix @cirvixai]

We built Cirvix because "trust the model" is not a security policy.

[Cirvix @cirvixai]

Your agent can try anything. With Cirvix, it can only do what you allow.

[Cirvix @cirvixai]

Hot take: the agent isn't the risk. Its permissions are.

[Cirvix @cirvixai]

Every MCP server you install is a stranger with a key to your laptop.

[Cirvix @cirvixai]

git push --force should never be one token away from an LLM.

[Cirvix @cirvixai]

Approving every tool call is the same as approving none of them.

[Cirvix @cirvixai]

Prompt injection isn't a chatbot problem anymore. It's a shell problem.

[Cirvix @cirvixai]

A sandbox stops the agent from escaping. It doesn't stop the damage inside.

[Cirvix @cirvixai]

You review every PR from a junior dev. Your agent opens 40 a day. Who's reviewing those?

[Cirvix @cirvixai]

Your agent doesn't need to be evil. It just needs to read one bad README.

[Cirvix @cirvixai]

Be honest: what's the scariest command your AI agent ran without asking?

[Cirvix @cirvixai]

Claude Code doesn't need to read your .env. So why can it?

[Cirvix @cirvixai]

Unpopular opinion: auto-approve mode is a security incident on a timer.

[Cirvix @cirvixai]

"It's just a coding agent" is what everyone says right before the rm -rf.

[Cirvix @cirvixai]

Your AI agent has your AWS keys. It just doesn't know it yet.