MCP servers

You can give fleetd’s agents more tools through MCP servers: a GitHub or database server, your docs search, an internal API. Configure them in the desktop app, for every project or for one, and pick which agents get each.

A Navi’s Notion, and servers whose tools should ask before they write, are connected under its Integrations instead.

Add a server

Global servers are in Settings › Hub › Every project › Tools & skills, for the hub’s owners and admins only; a project’s own are in Settings › the project › Agents › MCP servers. Click Add server, then fill in:

Field	Is
Name	How agents see its tools: `mcp__<name>__<tool>`. Letters, digits, `_` and `-`. `fleetd` is taken by fleetd’s own tools.
How agents reach it	Command: a program each agent’s claude starts on its machine, speaking MCP over stdin and stdout, with its arguments one per line and its environment. HTTP (MCP’s streamable HTTP) or SSE (its older event-stream transport): a URL, with headers.
Environment or headers	Each value plain or secret. See Secrets.
Who gets it	Lead, Writers (write tasks’ agents), Reviewers (review and read tasks’ agents). A thread you start yourself gets none.
On	Off keeps it saved but gives it to no one.

A project’s server replaces a global one of the same name in that project, for every role, even while the project’s is off: that’s how to turn a global server off for one project.

A command server starts in the agent’s working directory with the agent’s environment, the project’s variables included, plus its own environment.

A Navi with a Google account connected also gets fleetd’s own google server, for its lead only: see Google for Navi. It replaces a server of the same name in that Navi.

When a change takes effect

An agent started after you save gets the change. An agent already running keeps the servers its claude started with; fleetd hands it the change with its next message, and it takes effect the next time that agent’s process starts (after it was idle, stopped or the node restarted). Removing a server works the same way.

Secrets

Mark a value Secret for a token, a password or an Authorization header. It is sealed in the app to each machine’s key, as a project’s secret variables are: a project server’s to its machine, a global one’s to every machine. The hub passes the sealed values on to the nodes without storing or logging them, and keeps only each node’s fingerprint of the value. The node keeps them in ~/.fleetd/node/mcp-secrets/ (mode 0600) and puts them into the MCP config it writes for each agent’s claude (also 0600). If an agent prints one, the node masks it as ***<server>.<NAME>*** before the output leaves the machine, as it masks a project’s secrets.

A secret is never shown again. Editing a server keeps a saved secret you leave blank; type a value to replace it.

Test connection

Test connection has the node start the server (or connect to it), with its secrets, exactly as an agent’s claude would, and list its tools within 30 seconds. A project’s server is tested on its machine in the project’s checkout; a global one on the first machine online. On failure it says why: a command that isn’t on the machine’s PATH, the command’s error output (secrets masked), an HTTP status such as 401 when a header is wrong, or no answer in time.

Your own MCP servers

Agents also get the MCP servers of your own claude setup (~/.claude.json, a repo’s .mcp.json), as before. Turn Agents also get their own MCP servers off in Settings › Hub › Every project › Tools & skills to give them only fleetd’s tools and the servers configured here: claude then runs with --strict-mcp-config. fleetd’s own tools (report, ask_lead, the lead’s tools) are always there and can’t be replaced.

From an MCP client

mcp_servers_get lists the servers, mcp_server_set adds or replaces one (each secret sealed to each machine’s boxPubKey), mcp_server_remove removes one, mcp_settings_set turns your own servers on or off, and mcp_server_test tests one. All but mcp_servers_get take a human key.