MiMo Code: Run Xiaomi’s Free Claude Code Alternative in 10 Minutes

Xiaomi released MiMo Code, an open-source AI coding agent that forks OpenCode and adds persistent cross-session memory, three days ago. The tool, which installs in one command and is available under an MIT license, has become the most-watched open-source AI coding agent on GitHub. MiMo Code offers three configuration paths, including a free anonymous mode and a paid tier at $1 per million input tokens, and claims to outperform Claude Code on tasks exceeding 200 execution steps.

MiMo Code dropped three days ago and is already the most-watched open-source AI coding agent on GitHub. Xiaomi’s terminal agent forks from OpenCode https://opencode.dev , adds persistent cross-session memory that Claude Code still lacks, ships under an MIT license, and installs in one command. If you’ve been eyeing alternatives to Claude Code — especially after this week’s agentic billing shock https://byteiota.com/github-copilot-billing-agentic-cost-jump-2026/ — here’s how to get MiMo Code running and when it’s actually worth using. Install in One Line Requirements: Node.js 18 or later. That’s it. Mac and Linux: curl -fsSL https://mimo.xiaomi.com/install | bash Windows via npm : npm install -g @mimo-ai/cli Run mimo after installation and a TUI setup wizard launches automatically. It walks you through three configuration paths depending on what you need. Three Ways to Configure It The setup wizard offers these options: MiMo Auto — Anonymous, zero-config mode using a shared inference pool. Rate-limited but free. Good for evaluating MiMo Code before committing to an API key. Not suitable for production or sustained work. Xiaomi MiMo Platform API Key — Sign up at platform.xiaomimimo.com https://platform.xiaomimimo.com , generate a key, and you’re on the paid-but-cheap tier: $1 per million input tokens and $3 per million output tokens as of the May 2026 permanent price cut. The free Flash model on OpenRouter https://openrouter.ai is also an option for lighter development work. Import Claude Code Auth — The path most existing Claude Code users will take. One step in the wizard imports your existing Anthropic credentials and routes them through MiMo Code’s harness. Before you do this, clear any conflicting environment variables: unset ANTHROPIC AUTH TOKEN ANTHROPIC BASE URL mimo If you’d rather configure it manually, set these environment variables: ANTHROPIC BASE URL=https://api.xiaomimimo.com/anthropic ANTHROPIC AUTH TOKEN=<your-mimo-api-key MiMo Code also supports 75+ OpenAI-compatible providers — Ollama for local models, OpenRouter, or any custom endpoint — configured through the TUI. Your First Task: Compose Mode Once configured, run mimo in any repo directory. To kick off an agentic task without micro-managing every step, press Tab to activate Compose mode. Give the agent a rough description — “refactor the auth module to use the new middleware pattern” — and it plans, codes, tests, and reviews without waiting for approval at each step. This is functionally equivalent to Claude Code’s agentic mode. The key difference shows up when tasks run long. The Memory System: Why This Actually Matters This is the real reason to try MiMo Code, not the benchmarks. Claude Code loses everything at session end. Restart Claude Code tomorrow and it has no memory of today’s architectural decisions, preferred patterns, or project-specific context. You re-explain from scratch. For short tasks this is fine. For a multi-day refactoring project, it’s genuinely painful. MiMo Code addresses this architecturally with two commands: /dream — Run this at the end of a session or after a significant milestone. It scans recent session traces, extracts persistent knowledge into project memory, deduplicates, and removes stale entries. Think of it as the agent summarizing what it learned about your codebase into a long-term store. /distill — Run this after you’ve repeated the same manual workflow a few times. The agent identifies high-confidence patterns from recent work and auto-packages them as reusable skills or subagents. Tedious tasks you do repeatedly become automated. Xiaomi’s self-reported data claims MiMo Code outperforms Claude Code on tasks exceeding 200 execution steps, with a win rate above 65% and 40–60% fewer tokens used. Take those specific numbers with appropriate skepticism — they’re unverified, and the comparison is against Claude Sonnet 4.6, not Opus 4.8. But the underlying mechanism is real: persistent memory prevents goal drift on long-horizon tasks in a way that purely session-scoped agents can’t. VentureBeat’s deep-dive https://venturebeat.com/technology/xiaomis-new-open-source-agentic-ai-coding-harness-mimo-code-beats-claude-code-at-ultra-long-200-step-tasks/ on the architecture is worth reading for more context. Honest Assessment: When to Use It MiMo Code is three days old. The benchmarks are self-reported. The ecosystem is weeks, not years, from maturity. Here’s where it actually makes sense: Use MiMo Code for: Long, multi-session tasks where Claude Code context loss is a real problem. Cost-sensitive workloads — the free Flash model on OpenRouter covers substantial development work. Open-source or non-sensitive repos where you’re comfortable with a newer tool. Experimenting with custom model backends Ollama, local weights . Stick with Claude Code for: Enterprise workflows where audit trails and explicit approval gates matter. Tasks where Claude Opus 4.8’s reasoning quality is the bottleneck. Any situation where you need a proven track record over several months of production use. The practical play for most developers is running both: MiMo Code for long-horizon autonomous tasks and cost control, Claude Code for high-stakes reasoning work. The memory import path makes switching between them frictionless. Get Started The MiMo Code GitHub repository https://github.com/XiaomiMiMo/MiMo-Code has full documentation, open issues, and the growing MiMo-Skills library. If you want the model without the harness, platform.xiaomimimo.com https://platform.xiaomimimo.com has API access and pricing. MiMo-V2-Flash is available free on OpenRouter for developers who want to test before committing. For context on what MiMo Code claimed at launch https://byteiota.com/mimo-code-open-source-benchmarks/ , see our earlier coverage. Three days in, MiMo Code is already a serious option. It won’t replace Claude Code for everyone today. But for long tasks and tight budgets, it’s worth the ten minutes to find out.