MiMo Code: Free Open-Source AI Coding Agent with Persistent Memory

MiMoCode is Xiaomi MiMo’s open-source, terminal-native AI coding agent. It can inspect a repository, edit files, run shell commands, manage Git, connect to MCP servers, and retain project knowledge across sessions.

The MiMo-V2.5 anonymous free trial ended on July 26, 2026. The MiMoCode CLI remains available under the MIT License and separate use restrictions, but model access now depends on the provider you connect.

Current model options include the Xiaomi MiMo platform, Codex OAuth for eligible ChatGPT Plus or Pro accounts, imported Claude Code authentication, and custom OpenAI-compatible endpoints.

MiMoCode is good for developers who want a provider-flexible coding agent with persistent memory, checkpoint-based context reconstruction, MCP, reusable skills, and deterministic multi-agent workflows.

Its most distinctive value appears in long-running repository work that spans several sessions.

MiMoCode Current Status

Free-status note: MiMoCode’s source code remains open source. The discontinued MiMo Auto trial referred to hosted model access, not the software license. Xiaomi-hosted model usage now requires an eligible account, balance, or Token Plan. Provider and API charges vary outside Xiaomi’s platform.

Key Features

• Repository-level coding: MiMoCode reads and writes files, runs shell commands, works with Git, searches code through LSP integrations, and supports interactive or headless execution.
• Build, Plan, and Compose agents: Build has full development tools, Plan provides read-only analysis, and the dedicated Compose agent supplies the legacy interactive orchestration path.
• Recommended Compose workflow: Current releases recommend the Build agent with the /compose-next skill for interactive spec-to-ship work. The dedicated Compose agent remains available for the legacy skill-driven workflow.
• Persistent project memory: MEMORY.md stores durable project knowledge and architecture decisions. Session checkpoints, scratch notes, task records, and SQLite FTS5 preserve working context.
• Context reconstruction: When the active model approaches its usable context limit, MiMoCode rebuilds the session from the latest checkpoint, project memory, task progress, and recent messages.
• Configurable compaction point: The /context-limit command can lower the context threshold for a model. This helps control latency, route-specific model limits, and high-context pricing tiers.
• Goal-verified autonomous work: The /goal command defines a natural-language stop condition. A separate judge checks whether the condition has been met before the agent finishes.
• Dream and Distill: /dream consolidates durable knowledge into memory. /distill turns repeated procedures into reusable skills, subagents, or commands.
• Subagents: Primary agents can launch general or exploration subagents for parallel and background work, with lifecycle tracking and cancellation.
• Deterministic workflows: JavaScript workflow scripts can encode fixed phases, bounded retries, structured handoffs, parallel execution, nested workflows, and isolated Git worktrees.
• Four built-in workflows: MiMoCode includes compose, deep-research, fact-check, and research-experiment.
• Reusable skills: Built-in skills cover coding delegation, Codex, Claude Code, research, office documents, product design, modern Python tooling, long-horizon research, and other specialized tasks.
• Skill discovery: MiMoCode searches skills by exact name, localized alias, and BM25 relevance. A high-confidence match can load automatically, while /skills and slash autocomplete expose manual selection.
• Multi-skill orchestration: Referencing two or more skills in one message can load up to three skills and add an orchestration plan.
• Programmatic tool orchestration: The optional tool_script capability chains several tool calls inside a QuickJS sandbox to reduce repeated model round trips.
• Provider flexibility: First-launch options include Xiaomi MiMo, Codex OAuth, Claude Code authentication import, catalog providers, and custom OpenAI-compatible endpoints.
• Local model support: A local model can connect through an OpenAI-compatible gateway when the model and endpoint support the tool calls needed for agentic coding.
• MCP and development extensions: MiMoCode supports local and remote MCP servers, OAuth-enabled MCP authentication, LSP, plugins, custom commands, custom tools, hooks, formatters, and themes.
• Permission controls: Rules can allow, deny, or ask for actions. Destructive operations receive extra protection, while unattended modes retain prominent warnings.
• Voice input: Streaming voice input uses TenVAD and MiMo ASR. It requires sox and appropriate model access.
• Experimental Max Mode: MiMoCode can generate several candidate responses and use a judge to select one. This mode increases provider calls and token use.

MiMo Code vs Claude Code

MiMo Code and Claude Code both inspect repositories, edit files, run terminal commands, work with Git, automate development tasks, and connect to MCP servers. Their main difference is the model and extension architecture.

MiMoCode exposes its TypeScript source, inherits core OpenCode capabilities, and accepts several provider routes. Claude Code is Anthropic’s commercial coding agent and centers on Claude models through a Claude subscription, the Anthropic API, Amazon Bedrock, Google Vertex AI, or supported gateways.

Install MiMoCode

Install MiMoCode on macOS or Linux:

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

Use the native PowerShell installer on Windows:

powershell -ep Bypass -c "irm https://mimo.xiaomi.com/install.ps1 | iex"

The npm package works across supported platforms:

npm install -g @mimo-ai/cli --registry https://registry.npmjs.org

Open a project directory and run the mimo command:

cd /path/to/your/project
mimo

The first launch guides you through model access. Choose Xiaomi MiMo OAuth, Codex OAuth, Claude Code authentication import, a catalog provider, or a custom provider. The discontinued MiMo Auto trial no longer provides anonymous model inference.

Run /init inside the project on first use. MiMoCode analyzes the repository and creates or updates AGENTS.md with project structure, conventions, and guidance for later sessions.

/init

Select the primary agent before sending the first message. Build and Plan can switch after a session starts, but Compose is isolated once entered. Current releases recommend Build with /compose-next for interactive spec-to-ship work.

Start with a concrete repository task, review each requested permission, and inspect the resulting diff. Add a stop condition before a long autonomous run:

/goal all tests pass and the feature branch is committed

Use the recommended Compose skill for a feature that needs specification, implementation, verification, and review:

/compose-next implement the account export feature

Rebuild a long session from its latest checkpoint when context recovery needs manual intervention:

/rebuild

Choose an earlier compaction point for the active model when route limits, latency, or long-context pricing require a smaller working budget:

/context-limit

Run memory maintenance after a long project phase, or package recurring procedures into reusable instructions:

/dream
/distill

Model and Provider Options

A custom provider needs a registered model entry. Keep the upstream base URL and model ID exactly as the provider supplies them, and do not add or remove /v1 unless the endpoint requires it.

{
  "$schema": "https://mimo.xiaomi.com/mimocode/config.json",
  "model": "custom/MODEL_NAME",
  "provider": {
    "custom": {
      "name": "Custom",
      "npm": "@ai-sdk/openai-compatible",
      "only_configured_models": true,
      "models": {
        "MODEL_NAME": {
          "name": "MODEL_NAME"
        }
      },
      "options": {
        "baseURL": "BASE_URL",
        "apiKey": "API_KEY"
      }
    }
  }
}

MiMoCode Cheat Sheet and Quick Reference

CLI Commands

Run mimo <command> --help to inspect the current flags for a command. Notable TUI flags include --continue, --session, --model, --agent, --never-ask, --trust, and --dangerously-skip-permissions.

Application Slash Commands

Session and Prompt Commands

Built-in Prompt Commands

Built-in Agents

Compose Options

Built-in Workflows

Built-in Skills

The claude-code and codex skills appear only when the corresponding executables are installed. Project skills in .mimocode/skills/ and personal skills in supported skill directories can override a built-in skill with the same name.

Core Config Files

Core Config Fields

Important Environment Variables

Alternatives and Related Resources

• Best CLI AI Coding Agents compares terminal-based agents for developers who need a wider shortlist.
• Popular Agent Skills for Coding Agents lists reusable skills that extend compatible coding tools.
• Kimi Code CLI is another open-source terminal agent with skills and MCP support.
• Pi Coding Agent focuses on an extensible terminal workflow for developers.
• CodeWhale provides a DeepSeek-oriented terminal coding experience.
• OpenClaude provides a multi-model coding-agent CLI.

FAQs

Q: What is MiMoCode?
A: MiMoCode is Xiaomi MiMo’s open-source terminal AI coding agent. It can inspect repositories, edit files, run commands, manage Git, connect to MCP servers, and retain project knowledge across sessions.

Q: Is MiMoCode free?
A: The MiMoCode source code is available under the MIT License and separate use restrictions. The anonymous MiMo-V2.5 model trial ended on July 26, 2026. Continued model use requires Xiaomi MiMo access, an eligible Codex or Claude authentication route, another provider, or a custom endpoint. Provider charges can apply.

Q: What happened to MiMo Auto?
A: MiMo Auto supplied anonymous access to the MiMo-V2.5 trial. MiMoCode v0.1.9 enforced the end of that free trial after July 26, 2026. Bringing your own provider or API access remains available.

Q: Does MiMoCode require signup?
A: MiMoCode itself can be installed without creating a software-license account, but hosted model access requires the account or credentials used by the selected route. Xiaomi OAuth, Codex OAuth, Claude authentication import, and most catalog providers require an account.

Q: Can MiMoCode use ChatGPT Plus or Pro?
A: Yes. Current first-launch options include Codex OAuth for eligible ChatGPT Plus or Pro accounts. Availability and usage limits follow the Codex and ChatGPT account rules.

Q: Is MiMoCode an alternative to Claude Code?
A: Yes. Both tools can read code, edit files, run commands, work with Git, and connect to MCP servers. MiMoCode is open source and supports several provider routes. Claude Code is Anthropic’s commercial coding agent and centers on Claude models.

Q: Can MiMoCode use Ollama or another local model?
A: MiMoCode accepts custom OpenAI-compatible providers. A local model can connect when Ollama or another gateway exposes a compatible endpoint and the selected model can perform the structured tool calls required for coding work.

Q: Does MiMoCode support MCP servers?
A: Yes. MiMoCode can add and inspect local or remote MCP servers and supports OAuth-enabled MCP authentication. Permission rules should reflect the files, credentials, and actions exposed by each server.

Q: How does MiMoCode preserve context in long sessions?
A: A checkpoint writer records structured session state. Context reconstruction combines the latest checkpoint, project memory, task progress, notes, and recent messages. Budgeted injection limits how much stored material returns to the active context.

Q: What is the difference between /compose-next, the Compose workflow, and the Compose agent?
A: /compose-next is the recommended interactive path for exploratory feature work. The deterministic Compose workflow fits well-defined jobs that can run through fixed phases with limited intervention. The dedicated Compose agent is the legacy interactive path.

Q: Does MiMoCode have a desktop app or web interface?
A: The supported user interface is the terminal TUI, with headless and server commands for automation. The current command reference marks the separate web command as disabled, so this page does not treat MiMoCode as a finished desktop or browser coding product.

Q: Is permission skipping safe?
A: Permission skipping can let an agent run commands and access files with little or no confirmation. Use it only in a trusted disposable container, sandbox, or CI job. Keep explicit deny rules for sensitive paths and inspect the repository for malicious instructions.

Last updated: August 1, 2026.