OpenCognit
Guardrails

Risky commands
wait for your OK.

Agents run commands on their own. OpenCognit checks every one of them before it runs — a git reset --hard is refused, a git push waits for you, and nothing is ever waved through silently.

See it run

A push that waits,
until you allow it.

What you get

Protection for what agents
run on their own.

I

Three tiers

Block refuses the command, Approval holds it until you decide, Warning lets it run with a notice. Rules are regular expressions you can edit and extend.

II

Built-in rules

Blocked: git reset --hard, git clean -fdx, rm -rf on / or ~, DROP TABLE, mkfs. Held for approval: git push, npm publish, deployments, migrations, curl … | sh.

III

For all five CLIs

Claude Code, Codex, Gemini, Kimi and Hermes start with a small check hook — every shell command an agent runs on its own is asked about first.

IV

Approve in the tile

An approval card with Allow once, Allow for session and Deny appears right in the agent’s tile — without leaving the grid.

V

Unpacked first

Commands are unpacked before the check: bash -c, eval, $(…), chains, variables and disguises. Before git commit, the diff is scanned for keys and sensitive files.

VI

Never waved through

If OpenCognit can’t be reached, the agent asks you itself. Your own CLI configuration stays untouched — the hook only applies to sessions OpenCognit starts.

How it works

Three steps to safe agents.

I

Install OpenCognit

Linux or Windows. The built-in rules are active from the start.

II

Start your agents

Every session gets the check hook automatically.

III

Decide when it matters

Risky commands wait for you; adjust rules under Settings → Guardrails.

Questions about guardrails

What if I want to allow a push?

Click Allow once or Allow for session in the approval card. Deny sends the agent a refusal with the reason.

Can an agent work around a rule?

Commands are unpacked first (bash -c, eval, $(…), variables), and the agent gets a refusal with a request not to work around it.

Does it change my CLI settings?

No. The hook is added only to sessions OpenCognit starts; ~/.claude, ~/.codex and ~/.gemini stay untouched.

Can I add my own rules?

Yes. Every rule is a regular expression with a tier — add your own under Settings → Guardrails.

What does it cost?

Free with up to 4 agents and 2 projects. Pro is a one-time purchase — no subscription on top of your plan.