MiMo-Code: Xiaomi Answer to Claude Code, E Worth Am?

Wetin be MiMo-Code?

MiMo-Code na terminal coding agent wey Xiaomi MiMo team build, and dem release am as open source on 15 June 2026 under MIT license. E dey run as CLI (command line interface) tool for your shell. E dey read and edit files for project, run shell commands, control git, and keep notes about the codebase between sessions.

The structure dey similar to Claude Code. You go describe task for plain language, the agent go suggest edits and commands, and you go approve dem. Two differences matter. The source dey open, so you fit read exactly wetin the agent send to the model. And the model na separate choice from the agent, so the same harness fit connect to Xiaomi models, an OpenAI-compatible endpoint, or model wey you host yourself.

The repository name na MiMo-Code and the command wey e install na mimo. Both names dey for the documentation, so search for any one of dem go find the same tool.

MiMo-Code come from, and why e matter

MiMo-Code na fork of OpenCode, the open-source terminal agent from the sst project. The LICENSE file still get the original Copyright (c) 2025 opencode line, and the README get one short section about the relationship. One open issue dey for the repository, asking Xiaomi to give the OpenCode contributors more complete credit than just keeping one license line.

If you know this, one thing wey otherwise go look impossible go clear. Project wey first release for June 2026 already ship polished TUI (text user interface), plugin system, provider abstraction wey fit talk to many model APIs, MCP (model context protocol) support, and LSP (language server protocol) integration. Dem no do all this work within six weeks. The fork bring am come. The new parts from Xiaomi na the memory layer and the connection to Xiaomi own models.

If you don already run OpenCode for VPS, the layout and keybindings go feel familiar, because na the same code dey underneath.

Wetin "models and agents co-evolve" really mean

The repository tagline na "Where Models and Agents Co-Evolve". This phrase get two separate claims, but na only one get published details to support am.

The first claim concern runtime, and documentation dey support am. MiMo-Code dey keep persistent memory store for each project, backed by SQLite full-text search, together with session state and task progress. Two slash commands dey control am. /dream dey collect durable knowledge from recent sessions and write am into project memory. /distill dey turn procedure wey repeat into reusable skill. Xiaomi own engineering write-up talk say these dey run on cycle of roughly one to four weeks, instead of after every session. So the agent behaviour for your repository dey change over time, while the model file remain exactly the same.

The second claim concern training, and you need take care here. Xiaomi describe the post-training of MiMo-V2.5 as 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 mean say dem train the model with multi-step tool-calling episodes instead of single question-and-answer pairs. That one na real, and na why the model fit maintain consistency across long tool sequences. But Xiaomi never publish training report wey show say dem optimise the model inside this specific harness, with the harness changing in response.

So understand "co-evolve" as statement of intention from team wey own both sides. The benefit wey you fit actually observe dey narrower, but e still get value: the same people write the tool schemas and system prompt and release the model, so dem fit each other. Claude Code no fit work that way because its harness na client of fixed API model. Anthropic dey train its models for agentic use too, but you no fit swap the model out from underneath the harness.

How one MiMo-Code session dey look for terminal

When you run mimo, e go open full-screen terminal interface. You go type task, agent go read files, e go propose change, and e go ask before e write or run anything. Apart from default build mode, e get plan mode and compose mode. E also get subagent system wey dey hand part of job to another agent run. E carry plenty built-in skills too. These na prepackaged procedures wey agent fit call by name.

You need know some details before you start.

• By default, approval dey happen for each action. Flag --dangerously-skip-permissions go remove the prompts. The name correct, so use am only when mistake no matter.
• E fit run without interactive terminal. mimo run dey take prompt and execute non-interactively, na wetin script or CI job need. Web interface dey available too.
• Configuration dey inside files wey you fit commit. Project dey read .mimocode/mimocode.jsonc, while global settings dey inside ~/.config/mimocode/mimocode.jsonc.
• Credentials no dey inside those files. Dem dey inside ~/.local/share/mimocode/auth.json, so keep that path out of git and out of any backup wey you go share.

Agent wey fit run shell commands na program wey you don give control of your shell. Give am im own machine and im own user. The safest way to test new one na disposable VM wey you fit destroy after the session, because bad command go then cost you rebuild instead of restore.

MiMo-Code really beat Claude Code?

Xiaomi launch material make one specific claim: if you run the same model inside both harnesses, MiMo-Code score higher. Na published figures be these.

[
  {
    "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
  }
]

For SWE-bench Pro, Xiaomi report 62 percent for MiMo-Code, compared with 57 percent for Claude Code. For Terminal Bench 2, the figures na 73 percent and 68 percent. The meaning be say the model remain the same, so difference of about five points come from the agent system, not from the model.

Treat these numbers as vendor numbers. Xiaomi run the evaluation and publish am. MiMo-Code no appear for the public SWE-bench or Terminal-Bench leaderboards, so no independent group don reproduce the difference. Xiaomi dey clear about the limit of its own result. The company write-up say these benchmarks measure one-shot problem solving for issues inside one repository. E also say that for its own testing, both agents perform about the same below roughly 200 execution steps. The claimed advantage show for long runs, but one-shot benchmark no cover that case.

Na this be the honest way to read am. Five-point difference for benchmark wey vendor run na reason to try the tool. E no be reason to move the work wey your team depend on.

Wetin MiMo-Code go cost to run

For launch, Xiaomi release one free channel wey dem call MiMo Auto, and you fit use am without registration. From the start, dem describe am as something wey go dey available for limited time. The release notes for version 0.1.9 on 24 July 2026 talk about the end of that free MiMo-V2.5 trial. No build workflow on top am.

Paid access to Xiaomi platform cheap compared with the frontier providers. As of August 2026, the published international price for MiMo-V2.5-Pro na $1.00 per million input tokens when cache miss happen, and $3.00 per million output tokens, with context up to 256K. Prices fit change, so read the current page before you plan budget.

The other option na to use your own hardware. Dem publish the weights for MiMo-V2.5 and MiMo-V2.5-Pro on Hugging Face under the MIT license, so nothing stop you from serving dem yourself. Make you judge the scale well. Model wey get one million token context window no be work for ordinary VPS, and you go need a VPS wey get GPU attached or dedicated machine. Smaller local models na different kind of work, and e dey closer to running Ollama on a VPS to self-host an LLM.

MiMo-Code fit connect to other vendors too. E fit migrate existing Claude Code login, connect OpenAI account, accept API key for another provider, or communicate with any OpenAI-compatible endpoint. Check the terms of consumer subscription before you connect am, because provider terms decide which clients fit use subscription plan. CLI from another vendor no always dey covered.

You suppose move comot from Claude Code?

If you already dey pay for Claude Code and e dey work, no. Version 0.1.9 for late July 2026, six weeks after the first release, na the honest signal. Version number wey be 0.1.x mean say breaking changes fit happen between releases, and configuration format still dey change. E dey okay for tool wey you dey evaluate, but e fit cause problem for tool wey team depend on.

MiMo-Code deserve serious consideration for four situations.

• Na you dey handle the bill. Xiaomi token prices dey far below frontier vendors, and the agent itself cost nothing.
• You need read the harness. Open source mean say you fit inspect the prompts and tool definitions. This matter for security review or audit.
• You want the agent and model to be separate purchases, so changing one no go force you change the other.
• The code no fit comot from your network. Open weights make self-hosted backend possible instead of negotiation.

Cursor and Copilot na different comparison. Both dey inside editor, while MiMo-Code dey inside terminal. This difference affect more of your workday than any benchmark score, and the comparison of Claude Code against Cursor, Codex and Copilot explain the trade-off well.

MiMo-Code dey run for VPS wey you already dey rent

Terminal agent suppose dey for server. E go hold session open for hours, e need stable network, and e suppose continue to work even when you close laptop lid. The normal way na the same one wey people dey use for other terminal agents: long-lived session for small server wey you fit reconnect to from anywhere, as dem describe for how to run coding agent inside tmux for VPS.

Give am dedicated unprivileged user, and point am to clone of the repository instead of your only copy. The access wey e get na the access of the account wey e dey run as. Na this control still dey work when prompt go wrong.

The project dey distribute itself for two ways: shell installer wey dem serve over HTTPS, and npm package wey dem name @mimo-ai/cli for any machine wey already get Node.js. For server wey matter to you, prefer the npm package. You fit read wetin dey inside the package before e run, and you no dey pipe remote script directly enter shell. For the bigger picture of other things wey deserve space for that box, the roundup of self-hosted AI agents na the map.

FAQ

MiMo-Code free na?

The agent free and open source under MIT license, so you fit read and modify am. The model wey dey behind am na the part wey dey cost money. Xiaomi release one limited-time free channel wey dem call MiMo Auto when dem launch am, and the release notes from 24 July 2026 talk say that trial don end, so plan for paid access. As of August 2026, the published international price for MiMo-V2.5-Pro na $1.00 per million input tokens when cache miss happen, and $3.00 per million output tokens. You fit also point MiMo-Code to any OpenAI-compatible endpoint and pay another vendor instead.

MiMo-Code na the same thing as OpenCode?

E start as fork of OpenCode and e still share the core. The terminal interface, plugin system, provider abstraction and MCP integration all come from that codebase, and dem preserve the OpenCode copyright line inside the repository LICENSE file. Xiaomi add persistent memory layer wey SQLite full-text search support, extra agent modes, and direct access to its own models. If you sabi OpenCode, you already sabi most of how MiMo-Code dey behave.

I fit use MiMo-Code with Claude or GPT models instead of Xiaomi own models?

Yes. The first-run setup offer to migrate existing Claude Code login, connect OpenAI account, or accept API key for another provider. You fit configure any OpenAI-compatible endpoint manually for ~/.config/mimocode/mimocode.jsonc. Check the terms of consumer subscription before you connect am, because provider terms control which clients fit use subscription plan.

MiMo-Code safe to run for my main machine?

Treat am like any agent wey fit run shell commands. By default, e dey ask for approval before each action. --dangerously-skip-permissions turns those prompts off, and na that one cause most accidents. Run am as dedicated unprivileged user, against clone of the repository, and if possible for machine wey you fit destroy and rebuild. The blast radius na the account wey e dey run as, so make that account get only small permissions.