Skip to content

mimo-code.md

Latest commit

 

History

History
236 lines (191 loc) · 12.6 KB

mimo-code.md

File metadata and controls

236 lines (191 loc) · 12.6 KB
Raw

MiMo Code

Xiaomi's terminal coding agent, run headless by the OpenAgents daemon.

  • Registry entry: registry/mimo-code.json
  • Adapter: packages/agent-connector/src/adapters/mimo-code.js (extends opencode.js), with its pure helpers in mimo-code-runtime.js
  • Project: https://github.com/XiaomiMiMo/MiMo-Code

What it is

MiMo Code is a fork of OpenCode. mimo run --format json streams OpenCode's JSON events (step_start, text, tool_use, step_finish, error) and --session resumes a session, so the adapter reuses OpenCodeAdapter for stream parsing and the failure taxonomy, and overrides how the CLI is found, configured, stopped and how its failures are read.

Checked against CLI 0.1.14 and 0.1.15 (Linux x64) — the only two releases the download server still serves. 0.1.14 is the registry's min_version.

Install

Xiaomi's own script, not npm:

Platform Command
macOS / Linux curl -fsSL https://mimo.xiaomi.com/install | bash
Windows irm https://mimo.xiaomi.com/install.ps1 | iex

Both install one native binary (Bun-compiled, ~135 MB):

Platform Location
macOS / Linux ~/.mimocode/bin/mimo
Windows %USERPROFILE%\.mimocode\bin\mimo.exe, or %MIMOCODE_INSTALL_DIR%\mimo.exe when that is set

The scripts reach PATH through a shell rc edit or the user Path value, which an already-running app never sees, so paths.js lists both directories. The binary is spawned directly; there is no wrapper.

The first command on a machine migrates MiMo's database and imports the user's Claude Code transcripts (~/.claude/projects) into it — about a minute for ~1,600 transcripts. Later runs rescan quickly.

Configuration

Two ways in: a Xiaomi-account sign-in, or an API key. All three fields are optional, because the sign-in needs none of them.

Variable Required Meaning
LLM_API_KEY no — blank runs on the sign-in A MiMo API key (https://platform.xiaomimimo.com), a MiMo Token Plan key, or a key for the endpoint below
LLM_BASE_URL no, prefilled https://api.xiaomimimo.com/v1. Token Plan: https://token-plan-cn.xiaomimimo.com/v1 (also -ams, -sgp). Any OpenAI-compatible endpoint works.
LLM_MODEL for a non-MiMo endpoint Model id, e.g. mimo-v2.6-pro; sent upstream as written, a gateway's vendor/model ids included. Blank on MiMo's endpoints: the first -pro model mimo models xiaomi lists

The workspace's model choice, when set, replaces LLM_MODEL.

  • A MiMo endpoint (any *.xiaomimimo.com host) runs on MiMo's built-in xiaomi provider with this key and base URL, which keeps what MiMo's catalog knows about each model (context window, reasoning round-trip). With no model set the adapter asks mimo models xiaomi once and runs the first plain -pro model it lists (see below). A model that is set is declared status: "beta", because MiMo drops every model its catalog marks deprecated — the v2.5 line now — even when the run names it, while the key's API still serves them and the picker lists them from it.
  • Any other endpoint is declared for the run as an @ai-sdk/openai-compatible provider named openagents; a model is required.

Xiaomi-account sign-in

With no key, the run is on the Xiaomi-account sign-in. The adapter reads the xiaomi entry of auth.json (its key, and its Token Plan base_url if any) and declares it for the run exactly as it declares a form key, on the xiaomi provider. XIAOMI_API_KEY is not a way in: it is removed from the CLI's environment, and a signed-in agent's core environment drops it too.

What MiMo will run is its catalog minus what it marks deprecated: on 0.1.15 that is mimo-v2.6-flash, mimo-v2.6-pro and mimo-v2.6-pro-ultraspeed (signed in or not — it is not per account), and mimo-v2.5-pro is Model not found unless declared beta (see above). So there is no fixed default: with no model the adapter asks mimo models xiaomi once per agent and pins the first plain -pro it lists. Leaving the pick to MiMo instead hands it to the user's own model / agent.*.model settings, which name other providers and fail. The sign-in tab lists the same command (ids written xiaomi/<model>) and preselects the same model.

Key and sign-in stay apart

  • Each run's credential is declared in its own config — the form key, or the sign-in's key read from auth.json — with headers cleared, and the run starts with MIMOCODE_AUTH_CONTENT={}, which the CLI reads in place of auth.json. So nothing MiMo has stored or the user configured in MiMo (a provider.xiaomi.options.apiKey, an XIAOMI_API_KEY, the sign-in under a key run) stands in for it. A wrong form key fails auth_failed on a signed-in machine; a sign-in runs even with a wrong apiKey in the user's MiMo config. Only a model on a provider of the user's own MiMo setup (provider/model) reads auth.json, as theirs.
  • A sign-in run gets no form key: an agent set up on the sign-in tab carries OPENAGENTS_AUTH_MODE=cli_login, and the core drops every key and base URL (and XIAOMI_API_KEY) from its environment, whatever the type has saved.
  • The launcher labels each by its own path (a key reads "API key" whether or not the CLI is signed in; a sign-in agent is judged on the sign-in only), and the model lists come from different places: the key's endpoint, or the CLI.
  • Neither writes to the other: a key run leaves auth.json as it was.
  • The launcher's Sign in counts only MiMo's Login successful: a failed login exits 0, and an earlier sign-in still on disk would satisfy the probe.

mimo providers login -p xiaomi is not an OAuth session. It opens https://platform.xiaomimimo.com/authorize?pk=… with an X25519 public key; after the user signs in, the platform issues the CLI an API key of its own (listed on the platform as mimo-code-cli-key-…) encrypted to that key, and the CLI stores it in ~/.local/share/mimocode/auth.json (Windows too, under the user profile):

{ "xiaomi": { "type": "api", "key": "…", "metadata": { "uid": "…", "base_url": "…" } } }

base_url is set for a Token Plan account and the CLI routes there itself. The key does not expire; deleting it on the platform signs the agent out, which a run reports as auth_failed.

The page reaches the CLI one of two ways: a redirect to a server the CLI starts on localhost, or a code shown on the page to paste back at Paste code here if prompted > (three tries; then Decryption failed, max retries reached and exit 0). Those two "Decryption failed" lines are in the language MiMo picks from the time zone, then LANG (Chinese in China); the prompt, Login successful and Login failed are always English. The CLI opens the browser only with xdg-open, and the URL it prints is always the paste-a-code one.

The launcher's Sign in runs that command over pipes with CI=1, which stops the CLI opening a browser, opens the printed URL itself — one tab on every platform — and takes the code in the app. Only Login successful counts as signed in: a failed login exits 0 too, and an earlier sign-in on disk would satisfy a probe. Sign-in is read from auth.json (a xiaomi entry holding a key), by the launcher and by the core's check_ready (creds_file + creds_key). Like the CLI, both look under $XDG_DATA_HOME/mimocode/ when that is set (creds_xdg_data_file; an older core reads only the default path).

How a message runs

mimo run --format json --dangerously-skip-permissions --dir <workdir> --model <provider>/<model>
         [--title "OpenAgents - <channel>" | --session <id>] [--agent plan]

Environment:

Variable Why
MIMOCODE_CONFIG_CONTENT The run's config (below). MiMo loads it after the user's global and project config and removes it from the environment of the commands its tools run
MIMOCODE_EXPERIMENTAL_CRON=0 A one-shot run cannot fire a schedule; with the scheduler on, MiMo creates .mimocode/ in the project and the next run installs its plugin package there
MIMOCODE_DISABLE_AUTOUPDATE=1 A run does not replace the binary

LLM_API_KEY itself is removed from the CLI's environment: the key travels in the config, where the model's shell commands cannot print it.

The config:

  • provider — the key and base URL as configured (key mode only).
  • enabled_providers: [<that provider>], disabled_providers: [], and model, small_model, model_groups.{lite,standard,ultra} and agent.{general,explore}.model all set to the run's model. Without this the user's own MiMo config reroutes parts of a run under the user's own keys: a small_model received the first message of every session for titling, and a subagent ran on the user's agent.general.model. Providers MiMo picks up from the environment (DASHSCOPE_API_KEY, XIAOMI_API_KEY, …) or auth.json are likewise switched off. test/mimo-code-cli.test.js reproduces both against a real CLI.
  • mcp.openagents-workspace — the workspace MCP server (agent-connector mcp-server), its token in the server's own environment. MiMo names its tools openagents-workspace_<tool>.
  • instructions: [<briefing file>] — the workspace briefing, rewritten before every run under ~/.openagents/agents/<agent>/mimo-code/. MiMo puts it in the system prompt each turn, so mode changes and pinned decisions are current; it holds no key or token. MiMo concatenates instructions with the user's own.

Nothing is written into the project or into the user's MiMo config.

  • The prompt goes on stdin. run reads stdin to EOF whenever it is not a TTY, so it is always closed.
  • --dangerously-skip-permissions: a permission MiMo would ask for (writing outside the project, for one) is otherwise refused on the spot and the turn ends with no reply. The daemon is the approval boundary, as for every other agent here. Plan mode uses MiMo's plan agent, whose edit ban is a hard permission the bypass does not lift.
  • --title on a new session saves a request: without it MiMo asks the model to title the session, sending it the first message a second time.
  • The session id is stored per channel as soon as the stream names it, so a stopped turn's session is resumed by the next message.

Stop

MiMo has no signal handling: SIGINT or SIGTERM ends it at once. Each shell command its tools start runs in a session of its own, outside the CLI's process group, so signalling the group leaves those commands running. Stop therefore reads the process tree first (ps -A -o pid=,ppid=,pgid=), then signals the CLI's group and every descendant's group — SIGTERM, then SIGKILL after 1.5 s. On Windows taskkill /T /F is run straight away, while the CLI is still alive to walk the tree from.

A stopped session resumes; its first event is the interrupted tool call with status error.

Failures

Nearly everything exits 0:

Situation What the CLI does Category
Key rejected (401) error event on stdout auth_failed
Model MiMo's API does not serve HTTP 400 Unsupported model <id>, an error event at once model_not_found
Unknown model on another endpoint (404) retries for about three minutes, then an error event model_not_found
Server error (5xx) retries without end, nothing on stdout timeout (idle watchdog, 5 min)
No credential stderr error: MiMo free API service has ended. Sign in or configure a third-party API. credential_missing
--session it does not have stderr NotFoundError … Session not found: <id> session_missing — the adapter drops the session and runs the turn again in a new one

On stderr the error line follows a chunk of the CLI's bundled source and is followed by a stack trace; only the error line is classified.

Not yet verified

  • A Token Plan key. A MiMo API key on api.xiaomimimo.com/v1 with mimo-v2.6-pro was run end to end (reply, multi-turn on one session, an MCP tool, Stop and continuing after it), as was a Model Studio key through the custom-provider route. GET /v1/models with the key lists what the model pickers offer — including MiMo's ASR/TTS models, which are not chat models.
  • Windows: install layout read from install.ps1; the Stop path there is taskkill /T, not exercised.
  • A Xiaomi-account sign-in of a plain (non-Token Plan) account. A Token Plan account was signed in for real on 0.1.15 — the authorize page, the pasted code, Login successful, the stored tp- key with its Token Plan base_url — and run end to end with no key on xiaomi/mimo-v2.6-pro (reply, multi-turn on one session, an MCP tool, Stop and continuing after it) and with no model (MiMo picked mimo-v2.6-flash). The core's check_ready and the launcher probe both read that auth.json as signed in.