MiMo-Code is Xiaomi’s terminal coding agent, published under the XiaomiMiMo organization and built as a fork of OpenCode. The public pitch is simple: keep the terminal-native coding workflow, add persistent project memory, add a few opinionated agent modes, and make first launch easy through MiMo Auto or a custom OpenAI-compatible provider.

That makes it different from another thin model wrapper. MiMo-Code is trying to be a long-running coding workspace. It reads and writes files, runs commands, manages Git, resumes from checkpoints, and keeps project notes in a local memory system. The interesting part is not that it can edit code. The interesting part is the bet that an agent can stop losing context every time the session changes.

The caution is just as important. This is a very new repository. It was created on June 10, 2026 and had 4,696 stars, 359 forks, and 310 open issues as of 2026-06. The star curve is extreme: the sampled history moved from 101 stars on June 10 to 4,301 on June 11, then 4,696 on June 12. That is launch heat, not proof of settled reliability.

What MiMo-Code Adds On Top Of OpenCode

MiMo-Code keeps the broad OpenCode shape: a terminal UI, provider choice, LSP and MCP support, plugin-style extension points, and a codebase organized as a TypeScript and Bun monorepo. Its README says the fork adds persistent memory, intelligent context management, subagent orchestration, goal-driven loops, compose workflows, voice input, and a dream or distill loop for turning repeated work into memory or reusable skills.

The memory design is the clearest product decision. MiMo-Code uses SQLite FTS5 full-text search and writes several project-local artifacts: MEMORY.md for persistent project knowledge, checkpoint.md for session state, notes.md for scratch notes, and per-task progress logs under tasks/<id>/progress.md. On resume, it injects memory and checkpoints under a token budget instead of handing the model a blank project.

The agent model is also opinionated. build is the default agent with development permissions. plan is a read-only analysis mode. compose is for specs-driven development and skill-driven workflows. The README says Tab switches between primary agents, while subagents are created when the system needs them.

Install

The README documents two install paths. Use the one-line installer if you want the vendor-provided setup path:

curl -fsSL https://mimo.xiaomi.com/install | bash

Or install the npm package globally:

npm install -g @mimo-ai/cli

First launch opens an automatic configuration flow. The supported options are MiMo Auto, Xiaomi MiMo Platform OAuth, import from Claude Code, and a custom provider entered through the TUI. For teams with existing model gateways, the custom provider path matters because the README says it accepts OpenAI-compatible APIs.

Quick Start Shape

MiMo-Code does not document a long manual bootstrap. The intended first run is install, open the TUI, choose an auth path, then let the assistant operate inside the current project. The development setup for contributors is separate:

bun install
bun run dev
bun turbo typecheck

The root package.json says tests should not be run from the repository root. That small detail is worth noting for contributors because this is a monorepo with workspaces under packages/*, packages/console/*, packages/sdk/js, and packages/slack.

When MiMo-Code Fits

MiMo-Code is most compelling if your problem is session continuity. If you already like terminal coding agents but keep losing task state, project conventions, and partial decisions between sessions, its memory and checkpoint system is the reason to try it.

It also fits users who want a hosted low-friction channel at first launch but do not want to be locked into it. MiMo Auto is advertised as free for a limited time, while custom OpenAI-compatible providers are available in the same setup flow. That split gives the project a consumer-friendly on-ramp without removing the bring-your-own-provider path.

It is less convincing if you need a calm, proven production tool today. The issue tracker is already noisy. Recent reports mention Claude Code import failures, auth credentials not persisting, WSL install problems, text paste problems, voice input not working, runaway log files, and safety concerns around destructive commands. A new agent can still be useful, but it should not be treated like a mature replacement for a workflow you depend on.

MiMo-Code Vs Alternatives

Tool Stars as of 2026-06 Language License Best fit
MiMo-Code 4,696 TypeScript MIT OpenCode fork with memory, MiMo Auto, compose workflows, and Xiaomi branding
OpenCode 172,678 TypeScript MIT Baseline open-source terminal coding agent with a larger existing community
Void 28,813 TypeScript Apache-2.0 Open-source AI editor experience built around VS Code-style workflows
OpenClaw 378,109 TypeScript NOASSERTION Broad personal AI assistant positioning rather than a narrow terminal coding agent

The practical comparison is MiMo-Code versus OpenCode. If you want the upstream project with the largest visible user base, OpenCode is the safer default. If you want Xiaomi’s memory layer, MiMo Auto, compose mode, and the dream or distill workflow, MiMo-Code is the more experimental branch to watch.

Void is a different choice: editor-first rather than terminal-first. OpenClaw is broader still. It is closer to a personal assistant product than a direct MiMo-Code substitute.

Early Issues Worth Reading Before You Switch

The issue tracker is the part of the README that a launch page cannot give you. Several early reports are ordinary launch rough edges: WSL install errors, paste behavior in the input box, voice input failure, import from Claude Code failure, and a lingering opencode label during upgrade.

Two issue clusters deserve more caution. One report describes log files growing from hundreds of megabytes to many gigabytes, with comments mentioning WSL and macOS. Another describes an agent cleanup action that removed global npm packages without enough confirmation, followed by related concerns about high-risk commands. These are not abstract concerns for a coding agent. A tool that can run shell commands needs conservative permission boundaries, clear previews, and recoverable actions.

That does not make MiMo-Code unusable. It does mean the right test is a disposable repository or a project with clean backups, not your only copy of an important codebase.

Star Growth

The curve is almost vertical because the repository is only two days old. The useful reading is not “adoption is proven.” It is that a Xiaomi-branded OpenCode fork with memory became visible immediately, at the same time GitHub’s trending board is already full of agent skills and coding-agent utilities.

Read anomalyco/opencode first if you want the upstream reference point. voideditor/void is the editor-side alternative, while openclaw/openclaw shows a broader personal assistant direction. The same week also put infrastructure like apple/container next to agent tools, which is useful context for why a terminal agent can trend beside low-level developer tooling.

FAQ

What is MiMo-Code?

MiMo-Code is a terminal-native AI coding assistant from XiaomiMiMo. It is a fork of OpenCode that adds persistent memory, checkpoint-based context reconstruction, subagents, compose workflows, voice input, and Xiaomi MiMo service integration.

How do I install MiMo-Code?

The README lists two options: curl -fsSL https://mimo.xiaomi.com/install | bash or npm install -g @mimo-ai/cli. First launch then guides provider setup.

Does MiMo-Code require Xiaomi’s model service?

No, not according to the README. MiMo Auto and Xiaomi MiMo Platform OAuth are supported, but the setup flow also supports custom OpenAI-compatible API providers.

Is MiMo-Code the same as OpenCode?

No. It is built as a fork of OpenCode and keeps core OpenCode capabilities, but adds Xiaomi’s memory, context, subagent, goal, compose, voice, dream, and distill features.

Is MiMo-Code stable enough for daily work?

Treat it as early-stage. As of 2026-06 the repository was two days old and had hundreds of open issues. Test it in a backed-up or disposable project before letting it run broad shell or package-management changes.

What are the main MiMo-Code issues right now?

Recent reports mention auth persistence, import from Claude Code, WSL setup, paste handling, voice input, log growth, and command-safety concerns. Those areas matter more than cosmetic polish because they affect trust in an autonomous coding tool.