# Claros

> Your agents know the tools. They don't know how you work.

Claros keeps a team's unwritten rules, taste (brand voice, design system, conventions) and compliance (approval limits, money, data), each with a named owner, and checks people and their AI agents at the moment of action. A broken rule gets one of four responses: stop, ask the owner, nudge with the fix, or watch (logged, not enforced). It is not tied to one app or one team.

## Why

Agents know the app: how to fill a form or call an API is in their training data and the app's docs. They don't know your team's own rules, which live in people's heads and differ at every company. When an agent breaks one, it usually surfaces weeks later, in a review or an audit.

## How it works

1. Connect tools and documents. Connectors (Slack, GitHub, Google Drive, Notion, Linear, or their exports) and policies become suggested rules, each with an owner: who decided it. Or add a rule in one sentence on Home, the way you'd tell a teammate; or, optionally, record rules by voice over your own screen.
2. The owner confirms. Mined and drafted rules act on nothing before that, then start in watch (counted, not enforced) until the owner turns them on; a drafted policy can skip watch with "Approve and enforce now". A rule you add yourself is live at once if it's just for you or you're an editor; a member's team rule runs in watch until an editor confirms it.
3. Checked at the moment of action: every write an agent makes, before it runs, and a person's Send or Submit with Claros for Mac. A stopped action never reaches the app; the agent or person gets the rule, the owner's words and who to ask.
4. When Claros can't tell, it asks the rule's owner (in the Mac notch, a Slack DM or email). The answer becomes a precedent: the same action is never asked twice, the judge reuses past decisions on similar cases when it checks new ones, and consistent answers propose a new rule version that starts in watch.
5. Every check, answer and rule version is kept in a hash-chained, signed log: `getclaros verify`, a SIEM stream, an AIUC-1 evidence pack.

## Where it sits

- Agents (live): `npx getclaros` signs you in through the browser and wires Claude Code, Cursor, Codex, Claude Desktop and VS Code: native hooks for shell and edit actions, and a local proxy in front of the app's MCP server (reads pass through, writes are checked). Browser agents: one line in Playwright (`await guard(page)`) or browser-use. Apps with no API: via Pomerado, each call checked the same way.
- People (preview): Claros for Mac, in the notch, checks commit moments (Send, Submit, Approve, ⌘↩) in any app. Text is redacted on the Mac; only the action is sent; nothing is recorded. Unsigned build.
- Company (preview): connectors mine suggested rules with owners; not yet run on a real workspace.

## Example (the demo on the home page)

The home page plays three short scripted scenes, all demo data with invented names:

- An AP agent books a 7,200 EUR CNC lathe to an expense account. Stopped; the invoice stays in Draft. The AP lead's words: "Anything that is equipment and over 5,000 is capex – you don't book it as an expense."
- A coding agent styles a button with a hand-picked hex value. The commit goes through with a nudge and a fix (`var(--primary)`), from the design systems lead's rule.
- A sales agent floats a 45% discount in a quote email. Held for the deal desk, from the sales director's rule.

Below that, three of the support lead's refund rules meet three actors: a tool-use agent stopped at the MCP guard, a person asked in the Mac notch as they press Send, and a browser agent nudged through Playwright.

## Privacy

From a Mac, only the action being checked leaves the device, redacted first. Connector text is sent to the model as written, to draft rules. Details: /privacy on this site.

## Contact

Get started: anyone can sign up at /signin. Teams can book a demo from the home page and talk to the founder, Timur Monasypov (operating as Claros, Massachusetts, USA). A password-gated preview of the app is at /preview. Index for agents: /llms.txt.