Skip to main content

Let coding agents use Crest

On Mac, coding agents you run, such as Claude Code or Codex, can use Crest as their browser: list and open tabs, read pages, click and type, take screenshots, and read a page's console and network requests. They work in the Spaces you choose, with the sites you're already signed in to there.

Agents are off until you turn them on.

Turn on agents​

  1. Open Settings > Agents.
  2. Turn on Allow agents to use Crest.
  3. Under Spaces, turn on each Space agents may use.
  4. Under Connect an Agent, choose Copy Setup for your agent, and paste it in your agent's terminal or settings.

Crest must be open while an agent works. Crest never opens a window for an agent.

Allow an agent​

The first time an agent uses Crest, Crest asks whether to allow it. The question names the program macOS verified by its code signature, not the name the agent gives itself. Crest remembers your answer and lists the agent under Allowed agents. Choose Forget to disconnect it; Crest asks again the next time it connects.

When the program is a general interpreter, such as node or python, allowing it covers any program it runs under the same agent name.

See what agents do​

  • Each tab an agent uses shows the agent's badge and an outline in its colour, and the page shows a bar naming the agent.
  • Choose Stop in that bar, or Stop Agent in This Tab from the badge, to keep the agent out of that tab, through any of its connections, until Crest quits or you forget the agent.
  • Settings > Agents lists the agents connected now. Choose Disconnect to end one's connection.
  • Turning agents off disconnects every agent.

What agents can reach​

  • Only the Spaces you turned on. A locked Space shows agents its name and nothing in it.
  • Never a private window, and never Crest's own pages, such as Settings, a browser or extension page, or a file.
  • Only http and https pages, and blank ones.
  • Not your saved passwords. A password field's value is left out of what agents read, cookies and sign-in headers are hidden in the requests they read, and request and response bodies are never shown.
  • Scripts in a page's own world only when you turn on Let agents run scripts in pages. A script runs with the page's own access, so it can read anything the page holds, including a password filled in on it.

Any program that runs as you on your Mac can run an agent you allowed. Crest's question tells agents apart so you can see and control them; it is not a defense against harmful software already on your Mac. Allow only agents you trust, and turn agents off when you don't need them.

For agent authors​

Crest includes a Model Context Protocol server as a helper inside the app, which an agent runs over its standard input and output:

/Applications/Crest.app/Contents/Helpers/crest-browser mcp

It speaks protocol version 2026-07-28 and the earlier versions that open with initialize. Call identify first with your name and brand colour as #RRGGBB, so the person sees which tabs you use. Then use list_spaces and list_tabs, and read pages with take_snapshot, whose element uids click, hover, fill and scroll take. The other tools are open_tab, navigate, show_tab, close_tab, screenshot, type_text, press_key, wait_for, list_console_messages, list_network_requests, get_network_request and evaluate_script.

When a page runs on WebKit, input reaches it as script events, which some pages ignore, and its network requests show their address, type and timing without their method, status or headers.