Every team runs on rules nobody wrote down: the refund limit, the deploy that needs a rollback plan, the discount the deal desk must see, the wording the brand never uses. Some are taste, some are compliance. Agents don’t know them, and most live in a few people’s heads. Claros keeps those rules, each with a named owner, and checks people and their AI agents at the moment of action: before an agent’s action runs, and at a person’s Send. A broken rule gets one of four responses: stop, ask the owner, nudge with the fix, or watch (logged, not enforced). When Claros can’t tell, it asks the rule’s owner, in the notch, a Slack DM or email (Asks & answers). The answer becomes a precedent: the same action is never asked twice, past decisions inform new checks, and consistent answers propose a rule version that starts in watch. Every check, answer and rule version lands in a signed, hash-chained evidence log.

How Claros works

Connect, confirm, check, learn.

Quickstart

Your first rule and your first stop in five minutes.

Guard your agents

npx getclaros: every write checked before it runs.

Claros for Mac

Checks at Send, Submit and Approve, with nothing recorded.

​
What makes it different

  • Every rule has an owner. Mined or proposed rules start as suggestions; a rule stops anything only after its owner or an editor confirms it.
  • Checked at the action, not before the task. The check runs when the agent calls the tool or the person hits Send.
  • It says when it can’t tell. A case the rules don’t settle becomes a question to the owner, never a silent pass.
  • Your agent doesn’t change. The guard sits in front of its tools; the agent doesn’t have to remember to ask.
On our benchmark, agents keep to a team’s rules more often with Claros: +16 pts accuracy over the team’s docs in the agent’s context, on both sets, and +24–29 pts over RAG on the same docs. Methods and caveats: Rules benchmark.
The guard is on npm (getclaros 0.4.1; 0.5.0, the one-command setup, isn’t published yet); the browser hooks are on npm (getclaros-browser) and PyPI (claros-browser). What works today and what’s next: the Roadmap.