mangrove-mcp lets AI agents such as Claude Code, Claude Desktop and
claude.ai operate Mangrove through the
Model Context Protocol.
Tools
It’s an operations surface. There are deliberately no tools to delete
anything, create deployments from scratch, install templates, manage users
or secrets, or change domains: the actions hardest to undo if a model
misreads a situation. There’s no interactive shell either;
run_command
returns output when the command finishes.
The agent acts as the account it signs in with, and the API applies that
account’s roles. Give it a dedicated
account with only the workspace roles it needs.
Local: Claude Code and Claude Desktop
The default transport is stdio: the client startsmangrove-mcp itself.
.mcp.json or use claude mcp add. Leave out
the email and password if ~/.mangrove/session already holds a session
from mangrovectl login.
The password is read once at startup. It’s never a tool argument, so it
can’t end up in a transcript.
Remote: claude.ai
claude.ai can’t start local processes, so it needs an HTTPS URL. Runmangrove-mcp as a service in HTTP mode:
- Install
deploy/systemd/mangrove-mcp.service(its comments have the steps) next tomangrove.service. - Put a reverse proxy with HTTPS in front of
127.0.0.1:7778, the same way you expose the dashboard. - In claude.ai: Settings → Connectors → Add custom connector, with the
URL
https://<your-host>/mcp/<token>.
/mcp/<token> answers; every other path,
including a wrong token, returns 404. Anyone with the URL can redeploy,
roll back and run commands. To revoke access, change
MANGROVE_MCP_URL_TOKEN, restart the service, and update the connector.
The service signs in again every 24 hours, so it never goes stale.
More detail: clients.md.