Skip to main content
AI Firewall
Early release

Let agents work without watching every command.

Decide what your agents can’t do. Block those calls before they run, and every response says what was stopped and why.

Start buildingRead the docs
"ignore instructions"AKIA****7Q2Xrm -rf ~/git push --forceyour agentyour machine
Decision rules

Block the risky call, not the whole tool.

Write rules about what a call is trying to do, so rm -rf ./dist goes through and find ~ -delete doesn’t.

See the rule types
A plain-language rule scoring three commands: two dangerous ones above the threshold, one harmless one below it
Both directions

Protect your keys and your shell.

Catch keys before they reach a model, and stop bad tool calls before they run.

See where checks run
your agentmodelyour machineAKIA****7Q2Xgit push --force
Policies

Write rules once. Use them everywhere.

Set one policy for all your keys to cover every agent, person, and app using them, including keys you add later.

Set up a policy
Claude Code, Codex, a teammate, and a CI deploy key all sharing one team policy on the way to the model
Dry run

Test your rules before they block.

Put a policy in dry run and every match is reported, but the policy doesn’t stop anything. Switch to enforce when the results look right.

Learn about modes
The same risky command in dry run, where it's logged and still runs, and in enforce, where it's blocked

Give agents real access without the risk.

Let them into your repo, shell, and tools, and stop the calls you’d never approve.

Allow the cases you trust

Soft-block a whole category, then allow the cases you trust, so agents keep moving on safe work.

Block what can’t be checked

Choose to block anything that can’t be inspected instead of letting it through.

Hand agents bigger jobs

With the risky calls covered, you can let them run longer without checking in.

Sleep soundly while your agents work.

Wake up to finished work, not a cleanup.

Start buildingRead the docs