MiMo-Code: Xiaomi's answer to Claude Code

What is MiMo-Code?

MiMo-Code is a terminal coding agent built by Xiaomi's MiMo team and released as open source on 15 June 2026 under the MIT license. It runs as a CLI (command line interface) tool in your shell. It reads and edits files in a project, runs shell commands, drives git, and keeps notes about the codebase between sessions.

The shape is the same as Claude Code. You describe a task in plain language, the agent proposes edits and commands, and you approve them. Two differences matter. The source is open, so you can read exactly what the agent sends to the model. And the model is a separate choice from the agent, so the same harness can point at Xiaomi's models, at an OpenAI-compatible endpoint, or at a model you host yourself.

The repository is called MiMo-Code and the command it installs is mimo. Both names appear in the documentation, so a search for either one finds the same tool.

Where MiMo-Code came from, and why that matters

MiMo-Code is a fork of OpenCode, the open-source terminal agent from the sst project. The LICENSE file keeps the original Copyright (c) 2025 opencode line, and the README carries a short section on the relationship. There is an open issue on the repository asking Xiaomi to credit the OpenCode contributors more fully than a preserved license line does.

Knowing this explains something that otherwise looks impossible. A project first released in June 2026 already ships a polished TUI (text user interface), a plugin system, a provider abstraction that speaks to many model APIs, MCP (model context protocol) support and LSP (language server protocol) integration. That work was not done in six weeks. It arrived with the fork. What is new from Xiaomi is the memory layer and the tie to Xiaomi's own models.

If you have already run OpenCode on a VPS, the layout and the keybindings will feel familiar, because it is the same code underneath.

What "models and agents co-evolve" actually means

The repository tagline is "Where Models and Agents Co-Evolve". That phrase carries two separate claims, and only one of them is backed by published detail.

The first claim is about runtime, and it is documented. MiMo-Code keeps a persistent memory store for each project, backed by SQLite full-text search, alongside session state and task progress. Two slash commands drive it. /dream pulls durable knowledge out of recent sessions and writes it into project memory. /distill turns a repeated procedure into a reusable skill. Xiaomi's own engineering write-up says these run on a cycle of roughly one to four weeks rather than after every session. So the agent's behaviour on your repository changes over time while the model file stays identical.

The second claim is about training, and here you should be careful. Xiaomi describes the post-training of MiMo-V2.5 as a three-stage pipeline: supervised fine-tuning, then large-scale agentic reinforcement learning (RL), then on-policy distillation from several specialist teacher models into one student. Agentic RL means the model was trained on multi-step tool-calling episodes instead of single question-and-answer pairs. That is real, and it is why the model holds together across long tool sequences. What Xiaomi has not published is a training report showing the model optimised inside this specific harness, with the harness changing in response.

So read "co-evolve" as a statement of intent from a team that owns both halves. The benefit you can actually observe is narrower and still worth something: the same people write the tool schemas and the system prompt and ship the model, so they fit each other. Claude Code cannot work that way, because its harness is a client of a fixed API model. Anthropic trains its models for agentic use as well, but you cannot swap the model out from underneath the harness.

What a MiMo-Code session looks like in a terminal

Running mimo opens a full-screen terminal interface. You type a task, the agent reads files, proposes a change, and asks before it writes or runs anything. Beyond the default build mode it ships a plan mode and a compose mode, plus a subagent system that hands part of a job to a separate agent run. It also carries a large set of built-in skills, which are prepackaged procedures the agent can call by name.

A few details are worth knowing before you start.

• Approval is per action by default. The flag --dangerously-skip-permissions removes the prompts. The name is accurate, so use it only where a mistake costs nothing.
• It runs without an interactive terminal. mimo run takes a prompt and executes non-interactively, which is what a script or a CI job needs. A web interface is included as well.
• Configuration lives in files you can commit. A project reads .mimocode/mimocode.jsonc, and global settings live in ~/.config/mimocode/mimocode.jsonc.
• Credentials do not live in those files. They go in ~/.local/share/mimocode/auth.json, so keep that path out of git and out of any backup you share.

An agent that can run shell commands is a program you have handed your shell to. Give it its own machine and its own user. The safest way to try a new one is a disposable VM you can destroy after the session, because a bad command then costs you a rebuild instead of a restore.

Does MiMo-Code really beat Claude Code?

Xiaomi's launch material makes a specific claim: run the same model in both harnesses and MiMo-Code scores higher. These are the published figures.

[
  {
    "tool": "MiMo-Code",
    "swe_bench_pro_pct": 62,
    "terminal_bench_2_pct": 73
  },
  {
    "tool": "Claude Code",
    "swe_bench_pro_pct": 57,
    "terminal_bench_2_pct": 68
  }
]

On SWE-bench Pro, Xiaomi reports 62 percent for MiMo-Code against 57 percent for Claude Code. On Terminal Bench 2 the figures are 73 percent and 68 percent. The framing is that the model was held constant, so a gap of about five points comes from the agent system rather than from the model.

Treat those as vendor numbers. Xiaomi ran the evaluation and published it, and MiMo-Code does not appear on the public SWE-bench or Terminal-Bench leaderboards, so nobody independent has reproduced the gap. Xiaomi is unusually direct about the limit of its own result. The company's write-up says these benchmarks measure one-shot problem solving on single repository issues, and that in its own testing the two agents performed about the same below roughly 200 execution steps. The claimed advantage appears in long runs, which is the case a one-shot benchmark does not cover.

That is the honest reading. A five point gap on a vendor-run benchmark is a reason to try the tool. It is not a reason to move the work your team depends on.

What MiMo-Code costs to run

At launch Xiaomi shipped a free channel called MiMo Auto, usable without registration. It was described as limited-time from the start, and the release notes for version 0.1.9 on 24 July 2026 refer to the sunset of that free MiMo-V2.5 trial. Do not build a workflow on it.

Paid access to Xiaomi's platform is cheap next to the frontier providers. As of August 2026 the published international price for MiMo-V2.5-Pro is $1.00 per million input tokens on a cache miss and $3.00 per million output tokens, for context up to 256K. Prices move, so read the current page before you plan a budget.

The other route is your own hardware. The weights for MiMo-V2.5 and MiMo-V2.5-Pro are published on Hugging Face under the MIT license, so nothing stops you serving them yourself. Be realistic about scale. A model with a one million token context window is not a job for an ordinary VPS, and you are looking at a VPS with a GPU attached or a dedicated machine. Smaller local models are a different exercise, closer to running Ollama on a VPS to self-host an LLM.

MiMo-Code also connects to other vendors. It can migrate an existing Claude Code login, connect an OpenAI account, accept an API key for another provider, or talk to any OpenAI-compatible endpoint. Check the terms of a consumer subscription before you connect it, because provider terms decide which clients may use a subscription plan, and a CLI from another vendor is not always covered.

Should you move off Claude Code?

If you already pay for Claude Code and it works, no. Version 0.1.9 in late July 2026, six weeks after the first release, is the honest signal. A 0.1.x version number means breaking changes between releases and a configuration format that is still moving. That is fine for a tool you are evaluating and painful for one a team depends on.

MiMo-Code earns a serious look in four situations.

• The bill is your problem. Xiaomi's token prices sit far below the frontier vendors, and the agent itself costs nothing.
• You need to read the harness. Open source means you can inspect the prompts and the tool definitions, which matters for a security review or an audit.
• You want the agent and the model to be separate purchases, so that changing one does not force changing the other.
• The code cannot leave your network. Open weights make a self-hosted backend possible rather than a negotiation.

Cursor and Copilot are a different comparison. Both live inside an editor, and MiMo-Code lives in a terminal. That distinction shapes more of your working day than any benchmark score does, and the comparison of Claude Code against Cursor, Codex and Copilot covers the trade-off in detail.

Running MiMo-Code on a VPS you already rent

A terminal agent belongs on a server. It holds a session open for hours, it wants a stable network, and it should survive a closed laptop lid. The working pattern is the same one used for other terminal agents: a long-lived session on a small server that you reattach from anywhere, described in running a coding agent in tmux on a VPS.

Give it a dedicated unprivileged user and point it at a clone of the repository rather than your only copy. Its reach is the account it runs as, which is the one control that keeps working when a prompt goes wrong.

The project distributes itself two ways: a shell installer served over HTTPS, and an npm package named @mimo-ai/cli for any machine that already has Node.js. On a server you care about, prefer the npm package. You can read what the package contains before it runs, and you are not piping a remote script straight into a shell. For the wider picture of what else earns a place on that box, the roundup of self-hosted AI agents is the map.

FAQ

Is MiMo-Code free?

The agent is free and open source under the MIT license, so you can read and modify it. The model behind it is the part that costs money. Xiaomi shipped a limited-time free channel called MiMo Auto at launch, and the release notes from 24 July 2026 refer to that trial ending, so plan for paid access. As of August 2026 the published international price for MiMo-V2.5-Pro is $1.00 per million input tokens on a cache miss and $3.00 per million output tokens. You can also point MiMo-Code at any OpenAI-compatible endpoint and pay a different vendor instead.

Is MiMo-Code the same thing as OpenCode?

It began as a fork of OpenCode and still shares its core. The terminal interface, the plugin system, the provider abstraction and the MCP integration all came from that codebase, and the OpenCode copyright line is preserved in the repository LICENSE file. Xiaomi added a persistent memory layer backed by SQLite full-text search, extra agent modes, and direct access to its own models. If you know OpenCode, you already know most of how MiMo-Code behaves.

Can I use MiMo-Code with Claude or GPT models instead of Xiaomi's?

Yes. The first-run setup offers to migrate an existing Claude Code login, connect an OpenAI account, or accept an API key for another provider, and you can configure any OpenAI-compatible endpoint by hand in ~/.config/mimocode/mimocode.jsonc. Check the terms of a consumer subscription before you connect it, because provider terms control which clients may use a subscription plan.

Is MiMo-Code safe to run on my main machine?

Treat it like any agent that can run shell commands. It asks for approval before each action by default, and --dangerously-skip-permissions turns those prompts off, which is how most accidents happen. Run it as a dedicated unprivileged user, against a clone of the repository, and preferably on a machine you can destroy and rebuild. Its blast radius is the account it runs as, so keep that account small.