mcode-mcp

MCP server that delegates coding tasks to your locally installed MiniMax Code CLI (mcode, npm package @minimax-ai/code).

It wraps the real mcode binary instead of bundling its own copy of the agent, so every call inherits your MiniMax login, models and ~/.minimax config.

This is not the MiniMax M3 API provider in Pi, and not mmx-cli.

Sibling of pi-cli-mcp, qwen-cli-mcp, kimi-cli-mcp and grok-code-mcp — same architecture, same principles, MiniMax Code behind the wheel. Design: SPEC.md. Behaviour: CHANGELOG.md.

Verified against MiniMax Code 0.5.9.

Install

npx -y mcode-mcp         # no install
npm install -g mcode-mcp # or global

Requires Node ≥ 22 and MiniMax Code installed with its own installer.

The installer keeps MiniMax Code in ~/.minimax-code (or $MCODE_INSTALL_DIR) and exposes a launcher, <prefix>/bin/mcode, that follows the current release and pins a Node matching MiniMax Code's native SQLite ABI. The adapter finds and spawns that launcher by itself; there is nothing to configure.

Resolution order: MCODE_MCP_BIN → <prefix>/bin/mcode → mcode on PATH. Set MCODE_MCP_BIN only to use a different mcode. Never point it into <prefix>/releases/<version>/…: that path changes on every self-update.

Claude Code

claude mcp add-json mcode -s user '{
  "type": "stdio",
  "command": "npx",
  "args": ["-y", "mcode-mcp"],
  "timeout": 3600000
}'

Any other MCP client

{
  "mcpServers": {
    "mcode": { "command": "npx", "args": ["-y", "mcode-mcp"] }
  }
}

Keep the server name short (mcode): it becomes part of the tool names your model sees.

Related MCP server: Claude Code Bridge

Tools

mcode

Profiles

The account is the tool's name, not an argument. There is no profile parameter. For a profile called work you get mcode_work, mcode_work_reply, mcode_work_models, mcode_work_context and mcode_work_history; the untargeted mcode and friends stay on whatever the launch parameter selected.

mcode_work({ prompt: "…" })        // this task runs on the `work` account
mcode_personal({ prompt: "…" })    // and this one on `personal`, at the same time

That is what lets several accounts be live in one agent at once: they are separate tools, so a model chooses the account by choosing what to call, and the choice survives being retried, forwarded, or read back out of a transcript. An argument cannot do that — one tool is one account per call, and every call site has to carry the choice.

You will not see any of it unless a profile is in play. The per-profile tools and mcode_profiles appear only when the server has been pointed at a profile (MCODE_MCP_PROFILE / MINIMAX_PROFILE) or one exists on disk. On a machine with neither — which includes everyone on a MiniMax Code that has never heard of a profile — the tool list and every argument schema are exactly what they were in 0.2.0, so there is nothing new to be puzzled by. Signing a profile in makes the tools appear without restarting the server.

Asking for an account that does not exist is an unknown tool, not a run: a name only resolves to a real profile, so there is no path from a typo to a directory.

A profile is a separate MiniMax account with its own data directory — ~/.minimax for the default one, ~/.minimax-<name> for a named one — holding its own credentials, sessions and settings. It is mcode's own feature; this server passes it through with the same --profile <name> flag and resolves its own paths with the same arithmetic, so a run and the transcript read back afterwards can never disagree about which account they belong to.

mcode_profiles()                    // what exists, what is signed in, and the tools that reach each
mcode_work({ prompt: "…" })         // a task on the `work` account

Names must be 1–64 letters, numbers, dots, underscores or hyphens, starting and ending with a letter or number — a name becomes a directory segment under your home, so ../../etc is refused rather than sanitised.

A profile only gets tools if its tool names are usable: they must be unique and made of letters, numbers, _ and - (so no dots, and a name short enough for 64 characters). A profile called reply would collide with mcode_reply, and work_reply next to work would both claim mcode_work_reply; such a profile gets no tools rather than a tool that could run on the wrong account. It is still listed by mcode_profiles and can still be the server's default.

A profile argument is refused (when a profile is in play) rather than ignored: the account is the tool you called, and a caller that still passes the old argument would otherwise be billed to the default account without a word.

The default account has a tool when the server default is elsewhere. With MCODE_MCP_PROFILE=work, plain mcode runs on work, so the implicit account is reached as mcode_default, mcode_default_reply, and so on.

The server's default is MCODE_MCP_PROFILE, falling back to mcode's own MINIMAX_PROFILE if that is set and ours is not. A malformed value makes the server refuse to start instead of quietly falling back to the default account — running every task on the wrong account is worse than not running. The chosen profile is pinned into the environment of every mcode this server spawns, so an inherited MINIMAX_PROFILE can never send a run to an account other than the one this server reads from.

Two things to know before using one:

• A profile must have credentials first. Signing in is interactive and has no headless equivalent, so do it in a terminal with mcode login --profile <name> (or give it a Token Plan key with mcode provider set-minimax-key --profile <name>). Until then mcode_profiles reports signed out or not created, and a run against it cannot authenticate.

• MINIMAX_DATA_DIR / MAVIS_DATA_DIR outranks the profile, exactly as it does in mcode: with either set, every profile resolves to that one directory and they stop being separate accounts. mcode_profiles says so when it detects this. MCODE_MCP_MINIMAX_CONFIG is likewise a single path with no profile indirection, and combining it with a named profile is refused: the run would read the profile's own config.yaml, so the edit would succeed and the context window would not move.

An answer from a non-default profile carries [profile: <name>] in its prefix.

A session keeps the profile it started under, and it cannot be changed by a reply — a session id only names a conversation inside one account, so mcode_reply refuses a different profile rather than asking another account to load an id it never issued. mcode_history and mcode_context may read a session from another account, since those are reads; mcode_history reports the profile it read, and a cursor issued for one profile is refused in another, because a byte offset means nothing across accounts. Note that a context_window edit is a real write to that profile's config.yaml.

A profile name that has no directory is refused before anything starts, rather than letting mcode create an empty account for it.

context_window

mcode exposes no flag and no ACP config option for the context window: the only way to reach it is the minimaxModelContextLimits entry in your own ~/.minimax/config.yaml.

It works on print and is refused on acp. On print this server writes a private copy of the config and passes --config, and the window verifiably takes effect. On acp it does not: MiniMax Code builds a session selection from the advertised model value, and that selection carries a provider, model and variant but no context limit, so the turn falls back to the window in your config however this server writes that file. Measured on 0.5.8 — asking for 512000 with an explicit model, the session still reported 1,000,000. Since mcode acp has no --config, and the config path cannot be redirected without also moving the data directory, there is no way to make it work on that transport. Asking for it there is an error that says so, and nothing is written.

Set MCODE_MCP_ALLOW_ACP_CONTEXT_WINDOW=1 to edit the shared entry anyway: the server holds it for the whole run under a cross-process lock — which is judged stale by whether its owner process is alive, never by age — and rolls it back compare-and-swap, so a change by the TUI or by you wins, another instance's start will not disturb a run in flight, and a crash is repaired on the next start. Until minimax-code#384 lands it will have no effect.

mcode silently ignores a window the model does not advertise, so confirm the result with mcode_context. Only processes this call starts are affected, and a session that already exists keeps the window it was created with.

Not supported (honest error, not a fake flag): follow_up on mcode_send, effort, allowed_tools, system_prompt_append, permission=ask.

mcode({
  prompt: "Read and execute the prompt: /abs/path/prompt.md",
  cwd: "/abs/path/to/repo",
  model: "minimax/MiniMax-M3.1-Flash-Preview"
})

// later:
mcode_reply({
  session: "<id from the prefix>",
  prompt: "Now check the error paths of those call sites."
})

Default permission is full. Delegation is only useful when the delegate can act. Headless smart can stop with INTERACTION_NOT_AVAILABLE if Runtime asks a question. ACP session/cancel is how a running turn is aborted; there is no documented mid-turn follow-up queue, so mcode_send follow_up errors instead of pretending to queue.

What comes back

Only mcode's final result plus aggregate stats — never the transcript, thinking, tool arguments or raw stdout.

A timed-out or cancelled run still returns [session: <id>] and [session-key: mcode:<id>] when mcode named the session. Resume it with mcode_reply; do not treat the deadline as a lost session.

mcode_history

Read-only snapshot of the native transcript. Does not prompt mcode.

mcode_history({ session: "<id from the prefix>" })

Source: <dataDir>/v2/sessions/YYYY/MM/DD/<time>-session_<base64(id)>/messages.jsonl, where <dataDir> is MINIMAX_DATA_DIR, then MAVIS_DATA_DIR, then ~/.minimax[-<profile>] for the session's own profile. The page reports the profile it read. Items are user / assistant / tool / gap. Thinking and image binaries are omitted. Pass cursor from the previous page to continue; include_tools: false hides tool calls/results but still advances the cursor.

Source

git clone https://github.com/minmax/mcode-mcp.git
cd mcode-mcp
npm ci
npm test

Releasing

Releases are published by the Publish npm package workflow when a v* tag is pushed. It authenticates through npm trusted publishing (GitHub OIDC), so there is no NPM_TOKEN secret. The tag must equal v<version> from package.json.

npm version patch   # bumps package.json and creates the vX.Y.Z tag
git push --follow-tags