Skip to main content
The Deck MCP server gives any MCP client access to your Deck organization. It can create sources and tasks, run tasks, answer interactions, and debug failed runs from their screenshots and browser logs.
The server is hosted, so there’s nothing to install. You sign in with your Deck account when you connect, and the server acts as you, with the permissions of your role in the Console.
If you’re using a coding agent, npx @deckco/cli init sets up the Deck CLI and adds its local MCP server to your agent clients for you.

Connect it

Then run /mcp, select deck, and sign in.

Signing in

When you connect, your client opens the Deck sign-in page. Sign in, choose an organization if you belong to more than one, and approve access. Access is split into two scopes: The server only offers the tools your granted scopes allow. To switch organizations, reconnect the server in your client (in Claude Code, run /mcp and re-authenticate deck) and choose a different organization on the sign-in page. Reopening the connection without signing in again keeps the same organization.

What it provides

The server registers the tools agents use most. Most other API operations your scopes allow are reachable through execute.

Credentials

No tool accepts a password or other secret. To add a credential, the agent calls create_credential_link and gives you the URL. You, or whoever holds the login, enter the credential on a page hosted by Deck. The link expires within an hour. Nothing reports back when it’s used, so the agent checks list_credentials for the new credential. See Credentials.

Runs and deletes

run_task starts a run against the live source. Destructive operations, such as delete_task and reset_task, need confirm: true in the call. Without it, the server changes nothing and asks the agent to check with you first. Tools carry read-only and destructive annotations, so clients that ask before writes can prompt you.

Sensitive tasks

For sensitive tasks, the server doesn’t return screenshots or browser logs, matching the Console. Run status, output, and interactions are unaffected.

Run it locally

The Deck CLI includes a local version of the server for clients that run local MCP servers. It uses the CLI’s credentials, from deck login or an API key, instead of signing in through the browser. It has the same task, source, and run tools, but not create_credential_link, list_organizations, or switch_organization. Use the hosted server to collect credentials, or run deck credentials create --link in a terminal.
deck mcp install adds the server to Claude Code, Claude Desktop, Cursor, and Windsurf, and prints the entry to paste for Codex and VS Code. --client <id> picks the clients. To add it by hand, have your client run npx -y @deckco/cli mcp serve. For example, in Claude Code:
deck mcp install --read-only writes --read-only into each client’s config. To use --tools or --all-tools, add them to the client’s command by hand.

Docs-only server

To search the docs without a Deck account, connect the docs server instead. It’s read-only and can’t see your organization.
Connect it the same way as the Deck server, with deck-docs as the name and no sign-in. You can also add it from any docs page: open the context menu at the top of the page to copy the URL or add it to Cursor or VS Code. The docs server also exposes the Deck skill as an MCP resource. For a skill, a prompt, and raw docs files for coding agents, see LLM context.