MiMoCode
A terminal-native AI coding assistant with persistent memory, intelligent context management, and multi-agent orchestration — where models and agents co-evolve.
Evidence shows permission prompts (e.g., external directory permission) and user confirmation mechanisms, but full permission model details are not provided. Data flow transparency is limited; no detailed explanation of how data is processed. Sensitive data handling: API keys stored in plaintext, but docs advise users to protect file permissions. Dependency security: dependencies are locked and patched, but no vulnerability scanning evidence. External effects: server mode exists but requires user protection. Rollback: checkpoint system exists but no explicit rollback functionality. Source attribution: copyright notices present but publisher identity unverified.
Self-consistency: README and code structure are consistent, but some inconsistencies (e.g., repo name vs package name). Dependency availability: many dependencies, but no complete dependency list or verification. Failure messages: error handling exists but no specific examples.
Audience and scenarios: aimed at developers, multiple usage scenarios. Capability boundaries: permission system and sandbox description. Trigger precision: precise triggering mechanisms for skills and commands. Environment fit: supports multiple platforms, but no detailed system requirements.
Information architecture: clear documentation structure. Install notes: multiple installation methods. Naming stability: version control but no naming conventions. Examples and FAQ: examples provided but no FAQ. Known limitations: security threat model and known issues. License: MIT license. Versioning and changelog: no changelog provided. Maintenance responsibility: security reporting process but no clear maintainers.
Output usability: multiple output formats and tools. Marginal value: unique features like memory system and skills. Cost-benefit: free trial but no detailed pricing.
Claim traceability: documentation and code but no verification links. Cross-source corroboration: no external verification. Fact-inference separation: documentation mixes facts and inferences.
- API keys stored in plaintext; users must protect file permissions.
- Server mode requires user-configured authentication; otherwise unauthenticated access.
- Many dependencies; no vulnerability scanning evidence.
- Publisher identity unverified; use with caution.
What does this agent do, and when should you use it?
MiMoCode is Xiaomi's terminal-native AI coding assistant built as a fork of OpenCode. It retains core OpenCode features like multiple providers, TUI, LSP, MCP, and plugins, while adding persistent memory, intelligent context management, subagent orchestration, goal-driven autonomous loops, compose workflows, and self-improvement via dream/distill. It uses SQLite FTS5 full-text search for cross-session memory, and features a multi-agent system (build, plan, compose), subagent system, task tracking, and built-in workflows such as compose, deep-research, fact-check, and research-experiment. It provides a zero-config MiMo Auto channel (free for a limited time) and supports connecting to any mainstream LLM provider API. Licensed under MIT, it offers both personal and enterprise deployment paths.
MiMoCode is a terminal AI coding assistant that reads and writes code, runs commands, manages Git, and maintains a persistent memory system for project understanding across sessions. It is invoked via the mimo command, providing a TUI interface with three primary agents (build, plan, compose) switchable via Tab. It bundles various skills (e.g., arxiv, codex, deep-research) and workflows (compose, deep-research, fact-check, research-experiment) that can auto-parallelize tasks. It uses /dream and /distill to extract knowledge from sessions and create reusable skills, and /goal sets stopping conditions evaluated by an independent judge model. It supports voice input (TenVAD + MiMo ASR) and stores state via MySQL. Configuration is JSONC-based, with project-level and global files, and JSON Schema validation.
- Terminal developers using
mimofor daily coding who want the AI assistant to remember project context across sessions and auto-resume task progress. - Developers needing multi-step development pipelines (e.g., spec-to-code delivery), using the
/compose-nextskill or compose workflow to auto-parallelize independent tasks. - Users conducting deep research who can use the
deep-researchworkflow, running parallel subagents to collect cited multi-source findings. - Developers needing fact verification or adversarial fact-checking with the
fact-checkworkflow's three-juror adversarial voting. - Developers optimizing mechanically verifiable metrics via the
research-experimentworkflow's autonomous loop (baseline, iterate, keep/revert, audit). - Users wanting zero-config start with MiMo Auto (free for limited time) or BYO provider with OpenAI-compatible API.
What are this agent's strengths and limitations?
- Persistent memory across sessions (SQLite FTS5) automatically injects context on resume, avoiding re-learning.
- Built-in workflows (compose, deep-research, fact-check, research-experiment) enable unattended parallel automation.
- Fork of OpenCode retains multi-provider support, can connect to any OpenAI-compatible API, and offers MiMo Auto free-for-limited-time channel.
- Self-improvement mechanisms (/dream and /distill) extract knowledge from session history and create skills.
- Subagent parallelism, task tree tracking, and automatic checkpoints support complex projects.
- macOS default Terminal (Terminal.app) is unsupported; use iTerm2 or VS Code integrated terminal.
- WSL clipboard and Windows non-UTF-8 locale may have CJK garble, requiring extra setup.
- Voice input depends on MiMo platform (ASR model) and requires
sox; non-MiMo users may be limited. - TUI may lag over SSH; requires local rendering + remote server pattern.
- Requires Node.js and npm; non-command-line users may face a learning curve.
How do you install or deploy this agent?
One-line install (macOS/Linux): curl -fsSL https://mimo.xiaomi.com/install | bash; Windows PowerShell: powershell -ep Bypass -c "irm https://mimo.xiaomi.com/install.ps1 | iex"; or npm install globally: npm install -g @mimo-ai/cli. After installation, run mimo. The first launch guides through configuration, offering MiMo Auto (free for limited time), Xiaomi MiMo Platform OAuth, Codex (ChatGPT) OAuth, import from Claude Code, or custom OpenAI-compatible endpoints.
How do you use this agent?
Run mimo to start the TUI. On first launch, follow prompts to choose a provider (e.g., MiMo Auto for direct use, or configure API keys). In the TUI, press Tab to switch between build/plan/compose agents; type / to browse skills or /skill-name to invoke directly; use /goal to set stop conditions; use /dream and /distill for self-improvement; use /voice to enable voice input. Configuration files are at .mimocode/mimocode.jsonc or ~/.config/mimocode/mimocode.jsonc, where you can set models, permissions, MCP, etc. Run mimo models <provider> to see context windows.
How does this agent compare with similar options?
Comparable to Claude Code and Codex CLI, but MiMoCode offers more comprehensive memory, context management, and workflow systems, with multi-provider support, making it a viable alternative or complement.
FAQ
What is the MiMo Auto free channel?
How do I configure a custom model provider?
provider field in config with baseURL, apiKey, and model ID as custom:MODEL_NAME, supporting any OpenAI-compatible API.How does the session persist across restarts?
Can it run unattended in CI environments?
--dangerously-skip-permissions to auto-approve all actions, but use cautiously in trusted containers or sandboxes.How to extend skills or custom workflows?
.mimocode/skills/<name>/SKILL.md or place a .js file in .mimocode/workflows/ to define custom workflows, or override built-ins with same names.