MiMo Code

MiMo Code is Xiaomi MiMo’s open-source (MIT) coding agent for the terminal — a fork of OpenCode with a persistent project memory, in-session subagents (“actors”) and its own MiMo models. In NeuroSquad each MiMo agent is a card with the real mimo running inside.

Measured on MiMo Code 0.1.15.

Install

npm install -g @mimo-ai/cli

What works

• Exact status. A small NeuroSquad plugin inside MiMo relays its events: working, finished, or waiting for you — with the call it wants to make (MiMo Code needs your permission: bash_delete — rm -rf dist), the question it asks, or a command its bash tool hands to you. See Finished / needs your input.
• Arrows, live. NeuroSquad’s MCP server is in the card’s config, and MiMo re-reads its tools when they change — an arrow drawn mid-session works on the next turn. Installed MCP servers ask before every call.
• Canvas mode. bash, webfetch, websearch, codesearch and exec are denied in the card’s config. See Canvas mode.
• Dangerous mode without a restart, like Claude Code: while the switch is on, the plugin answers each permission ask through MiMo’s own API. See Dangerous mode.
• Sessions. After a restart the card continues the same conversation (mimo --session). If the session was deleted, the card starts a fresh one instead of leaving MiMo on a blank screen.
• Actors. MiMo’s actors run inside the same session; the card stays working while one runs, and their requests are marked as a subagent’s in Usage & costs, Run Stats and Agent Pulse.
• Long sessions. Context meter, Compact (MiMo’s /compact), the usage-limit chip with automatic “continue”, the prompt queue, the journal and hand-off.
• Models and providers. A lead agent can change the model with agent_set_model (the same session continues), or move the conversation to Claude Code and back. OpenRouter and your own servers work, the key kept out of every file.
• Plugins. Memory, Context7, Graphify and Caveman reach MiMo’s turns; the token saver works in dangerous mode.
• From your phone. The card works in remote access; its Stop key sends Escape.

Good to know

• Your config is never written. Each card has its own MIMOCODE_CONFIG file that MiMo merges over yours. The card starts MiMo with --trust, which skips the folder-trust question for that launch only.
• Stop a turn with Escape (pressed twice). Ctrl+C closes MiMo, so NeuroSquad never sends it.
• A slower start. With the app’s servers and plugin MiMo takes about 20 seconds to start; the card waits before typing anything.
• Two side requests aren’t counted. MiMo asks a model for each session’s title and, after each answer, for a next-prompt suggestion. It doesn’t store them, so Usage can’t count them.
• OpenRouter models must be in MiMo’s catalogue (openrouter/<slug>); custom providers list their own models.
• Analytics. MiMo sends usage metrics to Xiaomi by default. That is MiMo’s own setting and NeuroSquad leaves it alone; MIMOCODE_ENABLE_ANALYSIS=false turns it off.