A view is the set of installed modules that answer on an adapter. MCP and the JSON API each have their own view. Installing a program never makes it reachable on its own; you add it to a view:

--as takes mcp, api, or both, comma-separated. A module must be installed before it can be added. Views are stored with the environment’s configuration and take effect the next time the environment is served.

mim start launches one server for the environment and returns. For a container environment the server runs in a container named morloc-serve-<user>-<env>; on the native backend it is a background process. It answers on one port:

-p maps a host port to the server’s port and is written HOST:CONTAINER, 8005:8005 here. Without it, mim serves on 9000 when MCP is in the view (8080 otherwise) and moves to the next free port if that one is taken; it prints the port it chose.

By default the server listens on the host’s loopback address, 127.0.0.1. Only programs on the same machine can reach it, so it needs no password. Once a server is running, edits to the views apply only after you replace it with --force:

To reach the server from elsewhere, --expose binds it on every interface (0.0.0.0). An exposed server must be protected by a token, a shared secret that every client sends with each request:

mim never generates a token and never writes one to disk. You choose it, and pass it either as --auth-token <token> or, to keep it out of your shell history and process list, in MORLOC_MCP_TOKEN. Every request to /mcp, /call, and /discover must then carry the header Authorization: Bearer <token>; a request without it gets 401 and {"error":"unauthorized"}. /health stays open so monitoring works without the secret.

--allow-plaintext is required with --expose because the server speaks plain HTTP: the token and the data travel unencrypted. A token stops strangers who find the port; it does not stop anyone who can watch the network. For use over an untrusted network, keep the default loopback bind and reach the server through an SSH tunnel, or put a TLS proxy in front of it.

--allow-no-auth serves an exposed server with no token at all. It exists for a trusted private network or a gateway that already checks callers, and is a bad idea anywhere else.

/call/<module>/<command> takes the function’s arguments as a JSON array, in order, and returns the result in a JSON envelope:

A failed call returns {"status":"error","error":"…​"} with HTTP status 500. /discover lists every module in the API view with its commands, and /discover/smiles gives each command’s arguments, types, and docstrings, so a client can find out what to call without reading the source.

When the MCP view is not empty, mim start prints a client configuration entry on standard output, and its status messages on standard error. The entry has this shape:

Redirect it to a file to hand it to an MCP client that reads this format. Treat the file as a secret, since it contains the token, and keep it out of version control. For an exposed server, <host> is the machine’s hostname; replace it with an address your clients can resolve.

To add the server to Claude Code:

Each exported function becomes one MCP tool, named <module><command> (smilesmw), with the docstrings as its description and the argument types as its input schema. The protocol details — the handshake, how arguments map to tool properties, and what is not exposed — are in Model Context Protocol (MCP).

mim status lists the servers of all your environments, with what each serves (mcp+api here) and its URL, marked (token) when one is required. logs and stop act on the default environment, or the one named with --env.