# Claros

> Your agents know the app. They don't know your rules. Claros keeps a team's unwritten rules (taste and compliance), each with a named owner, from their tools, policies or said out loud, and checks people and their agents against them at the moment of action. When it can't tell, it asks the rule's owner. Any app, any team.

Anyone can sign up at https://claros.sh/signin. The way in for agents is one command: run `npx getclaros` in a terminal your user can see. It signs your user in through the browser (they can make an account and a workspace there) and enrolls this machine's agents: native hooks for Claude Code, Codex and Cursor (shell commands, file edits), a local proxy in front of each app's MCP server (reads pass through), and the Claros MCP server, so every write is checked without the agent having to ask. Run in a Playwright or browser-use project, it also installs the browser hook (`guard(page)`, `guard_tools(tools)`) with the project's own package manager.

Hosted assistants that can't run commands (Claude.ai, ChatGPT custom connectors) connect to the MCP server with an agent key (`clr_…`) your user makes in Claros under Account → Agent keys.

## Set up

- [Agent setup guide](https://claros.sh/agents.md): run the command, verify it, the hosted-assistant path, and the rule for checking actions
- [MCP server](https://timurs-macbook-pro.tail8a0be4.ts.net/mcp): Streamable HTTP; SSE at https://timurs-macbook-pro.tail8a0be4.ts.net/mcp/sse. Tools: check_action, get_decision, report_override, find_guardrails, get_workflow, quote

## Product

- [About Claros](https://claros.sh/llms.md): who it's for, how it works, where it sits, the home-page demo, privacy, contact

## Optional

- [Docs](https://claros-docs.vercel.app): concepts, supervising agents, access control, API reference
- [Privacy](https://claros.sh/privacy): what Claros records and what it never sees