metis
Wholiver · Wholiver/metis
Metis is a coding agent that searches, remembers, executes, and verifies code across the terminal and desktop, boosting AI/LLM coding performance.
Metis is a coding agent that boosts AI/LLM coding performance by 50%
296 stars31License: MITUpdated 17 days ago0 opens
Preview
▍
What you can do with it
- Ask Metis to explain the functionality of a given code repository.
- Use Metis to review a specific file or a `git diff` for code quality and issues.
- Integrate Metis into your development workflow for automated code verification.
Safety check
Safe
No risky patterns found in the files we scanned.
Automated scan of the repo files. Always review code you run.
From the README
English · 简体中文
A coding agent that searches, remembers, executes, and verifies across terminal and desktop.
Partners · Quick start · Benchmark & Comparison · Key features · Documentation
Partners
| Logo | Partner | Link |
|---|---|---|
| SiliconFlow (硅基流动) — China's leading independent ecosystem token supply platform, providing Metis with efficient and flexible model inference. | siliconflow.cn |
Quick start
Desktop
Standalone application with built-in Metis CLI and Server runtime (no Node.js required):
- macOS (Apple Silicon): Download
Metis-*-macos-arm64.dmgfrom the latest GitHub Release and drag to Applications. - Windows (x64): Download
Metis-*-win-x64-setup.exeor.zipfrom the latest GitHub Release.
CLI installation
Requires Node.js >=22.19.0.
npm install -g @wholiver_hu/metis
metis
Run in any repository or directory:
metis "Explain this repository"
metis @src/main.ts "Review this file"
git diff | metis -p "Review this diff"
Use /login for subscription providers or configure an API key. See Quickstart for the complete guide.
Benchmark & Comparison
Terminal-Bench 2.1 Benchmark Results
In a controlled benchmark run using the same model (DeepSeek V4 Flash), same 89 real-world tasks, identical budget, and environment:
| Agent Framework | Model | Benchmark | Solved (Accuracy) | Architecture & Harness Advantage |
|---|---|---|---|---|
| 🏆 Metis | DeepSeek V4 Flash | Terminal-Bench 2.1 (89 tasks) | 73 / 89 (82.02%) | ✅ Recursive 5-role agents + Self-learning adaptations + Plan/Build |
| OpenCode | DeepSeek V4 Flash | Terminal-Bench 2.1 (89 tasks) | 60 / 89 (67.42%) | ⚠️ Single-thread flat tool execution |
| 📈 Improvement | Same Model & Budget | Same Environment | +14.6% (+13 tasks) | 🚀 Harness, adaptations, and verification gates alone |
Feature Comparison Matrix
| Capability | Metis | Claude Code | OpenCode | Cursor / Cline |
|---|---|---|---|---|
| License & Pricing | ✅ MIT ($0 Free) | ❌ Proprietary | ✅ MIT ($0 Free) | ⚠️ Commercial |
| Model Freedom | ✅ Any Model / OrcaRouter | ❌ Anthropic Only | ✅ Multi-Provider | ⚠️ Limited / BYOK |
| User Interfaces | ✅ TUI + React Desktop | ⚠️ Terminal Only | ⚠️ Terminal Only | ⚠️ IDE Only |
| Workflow Mode | ✅ Plan ↔ Build Dual-Mode | ⚠️ Single Flow | ⚠️ Single Flow | ⚠️ Chat / Inline |
| Multi-Agent System | ✅ Recursive L0→L4 (5 Roles) | ⚠️ Flat Subagents | ⚠️ Basic | ❌ None |
| Self-Learning Adaptations | ✅ Filesystem Profile & Workflow Evolution | ❌ Ephemeral | ❌ Ephemeral | ⚠️ Code Embeddings |
| Verification Gates | ✅ Test Gates + Video Evidence | ⚠️ Manual Bash | ⚠️ Manual Bash | ⚠️ Basic Linter |
| Headless Benchmark | ✅ Python Adapter + Trace | ❌ None | ⚠️ Partial | ❌ None |
Key Features
- Plan & Build Dual Workflows — Safely investigate in read-only Plan mode, then execute approved plans with a live-updating checklist in Build mode.
- Dual Interface for Terminal & Desktop — Work directly in your terminal via the rich interactive TUI, or use the dedicated React/Vite Desktop workspace on macOS and Windows.
- Recursive Multi-Agent System — Native named agents (
coordinator,planner,implementer,reviewer,verifier) with L0→L4 recursive delegation and Git Worktree isolation. - Self-Learning Adaptations & Resumable Sessions — Project profiles, learned workflows, and custom checks adapt safely on disk across sessions with full rollback support.
- Extensible & Model-Agnostic — Use any LLM provider (OpenAI, Anthropic, DeepSeek, OrcaRouter, Gemini, Groq, Ollama, vLLM) and extend with TypeScript plugins, Agent Skills, and MCP.
- Benchmark-Grade Reliability — Automated verification gates, video evidence inspection, and full Terminal-Bench & Harbor evaluation readiness.
Documentation
| Topic | Guide |
|---|---|
| Install, authenticate, and start | Quickstart |
| Commands and terminal UI | Using Metis · TUI |
| Providers and custom models | Providers · Use SiliconFlow in Metis · Custom models · Custom providers |
| Multi-Agent System | Named Agents & Delegation |
| Benchmark & Evaluation | TerminalBench & Harbor |
| Sessions and compaction | Sessions · Compaction |
| Extensions, skills, and packages | Extensions · Skills · Packages |
| Prompts and interface customization | Prompt templates · Themes · Keybindings |
| Programmatic integration | SDK · RPC · JSON |
| Video inspection | Video tool |
| Security and configuration | Security · Settings |
| Platforms and isolation | Windows · Termux · tmux · Containers |
| Self-Learning & Adaptations | Autonomous Self-Learning |
See the documentation index for every guide.
Developer information
npm run build # Compile TypeScript and copy runtime assets
npm test # Run the Vitest suite
npm run clean # Remove compiled output
npm run build:binary # Build the standalone binary
npm --prefix desktop run dev # Start the React/Vite Desk