Tools#Any agent#Coding #Code review #Testing

CLI and desktop coding agent for terminal workflows

Metis is a coding agent for terminal and desktop use. It supports plan and build modes, named multi-agent delegation, persistent SQLite memory, and verification steps like tests and video evidence. It is packaged as a global CLI plus a standalone desktop app, with docs and adapters for providers, sessions, skills, prompts, and integrations.

138 stars•14 forks•TypeScript•Updated 7d ago

Who it's for

Builders who want a terminal or desktop agent that can plan, code, review, and verify work with saved context.

What it delivers

You can keep context across sessions, delegate work to subagents, and verify changes before you ship them.

What it does

Plan and build modes

Lets you investigate in read-only Plan mode and then execute approved work in Build mode with a live checklist.

Recursive multi-agent system

Uses named agents like coordinator, planner, implementer, reviewer, and verifier with delegation across levels.

Durable session memory

Stores project knowledge and decisions in SQLite so sessions can resume after restarts and compaction.

Terminal and desktop interfaces

Provides a TUI in the terminal and a React/Vite desktop workspace for macOS and Windows.

Verification gates

Adds test gates, video evidence inspection, and benchmark-oriented checks before work is accepted.

Model and extension support

Works with multiple model providers and can be extended with TypeScript plugins, skills, and MCP.

How to get it

1Requires Node.js >=22.19.0.
```
npm install -g @wholiver_hu/metis
metis
```

2Run in any repository or directory

metis "Explain this repository"
metis @src/main.ts "Review this file"
git diff | metis -p "Review this diff"

README

English · 简体中文

A coding agent that searches, remembers, executes, and verifies across terminal and desktop.

Quick start · Benchmark & Comparison · Key features · Documentation

Quick start

Desktop

Standalone application with built-in Metis CLI and Server runtime (no Node.js required):

macOS (Apple Silicon): Download Metis-*-macos-arm64.dmg from the latest GitHub Release and drag to Applications.
Windows (x64): Download Metis-*-win-x64-setup.exe or .zip from 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 + SQLite memory + 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, memory, 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
Durable Memory	✅ SQLite + Vector Search	❌ 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.
Durable Memory & Resumable Sessions — Project knowledge and decisions persist in SQLite across restarts, context compactions, and session forks.
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 · 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

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 Desktop app in development
npm --prefix desktop run build # Build the renderer and Electron artifact

The package exports the Node.js SDK from @wholiver_hu/metis and the RPC entry point from @wholiver_hu/metis/rpc-entry.

Contributing

Contributions are welcome. See CONTRIBUTING.md for development, Extension and Package integration, testing, and AI-assisted contribution guidance.

License

Distributed under the MIT License.

Files in the repo

Repository payload•25 top-level entries

.github— GitHub workflows and templates
adapters— Provider and benchmark adapters.
desktop— Electron desktop app with the built-in workspace and runtime.
docs— Documentation
eval_results— 1 item
examples— Worked examples
scripts— Helper scripts
src— Source code
test— Tests
vendor— 3 items
.env.example— Environment variables to set
.gitignore
AGENTS.md— Instructions for Codex and other agents
CHANGELOG.md
CONTRIBUTING.md
CONTRIBUTING.zh-CN.md
LICENSE— License
npm-shrinkwrap.json
package.json— Node package: install with npm
pnpm-workspace.yaml
README.md— The README
README.zh-CN.md
tsconfig.build.json
tsconfig.examples.json
vitest.config.ts

Discussion (0)

Ask about usage, or say what you built with it

No comments yet. Be the first to say what this is good for.

More tools

JuliusBrussee/

caveman

JuliusBrussee/cavemanTools

🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman

105k

1 add

MemPalace/

mempalace

MemPalace/mempalaceTools

The best-benchmarked open-source AI memory system. And it's free.

59k

stablyai/

orca

stablyai/orcaTools

Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.

66k

farion1231/cc-switchTools

A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode, OpenClaw, Grok Build & Hermes Agent. Only official website: ccswitch.io

132k

diegosouzapw/OmniRouteTools

Never stop coding. Free MIT AI gateway: one endpoint, 352 providers (150+ free), 1200+ models Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Quota-aware auto-fallback, RTK+Caveman compression saves 15-95% tokens, MCP/A2A, Desktop/PWA. Built by 550+ contributors

64k

headroomlabs-ai/

headroom

headroomlabs-ai/headroomTools

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

71k