Metis

Metis is an open-source coding agent built in TypeScript by Wholiver, distributed under the MIT License. It combines a rich terminal UI (TUI) with a React/Vite desktop application for macOS and Windows, and supports any LLM provider including OpenAI, Anthropic, DeepSeek, Gemini, Groq, Ollama, vLLM, and OrcaRouter. The project reached v1.1.9 as of August 2026 and is actively maintained on GitHub.

What It Is

Metis is an agentic coding assistant designed to search, remember, execute, and verify tasks across both terminal and desktop environments. Unlike single-threaded coding agents, Metis uses a recursive five-role multi-agent architecture (coordinator, planner, implementer, reviewer, verifier) with L0→L4 delegation and Git Worktree isolation. It persists project knowledge and session decisions in SQLite with vector search, making sessions resumable across restarts and context compactions.

Multi-Agent Architecture

The core differentiator in Metis is its recursive multi-agent system. Five named agents handle distinct responsibilities and can delegate tasks recursively up to four levels deep:

• Coordinator — top-level orchestration and task routing
• Planner — read-only investigation and plan generation
• Implementer — code execution and file changes
• Reviewer — code quality and correctness checks
• Verifier — automated test gates and video evidence inspection

Git Worktree isolation keeps parallel agent work from interfering with the main branch.

Plan & Build Dual Workflow

Metis separates investigation from execution through two explicit modes. In Plan mode, the agent operates read-only, exploring the codebase and generating an approved checklist. In Build mode, it executes the approved plan with a live-updating checklist and automated verification gates. This separation reduces accidental side effects and gives developers a clear review checkpoint before any changes are committed.

Benchmark Performance

The README reports a controlled benchmark run on Terminal-Bench 2.1 using 89 real-world tasks with the same model (DeepSeek V4 Flash) and identical budget across agents. According to the vendor-published results, Metis solved 73 of 89 tasks (82.02%), compared to 60/89 (67.42%) for OpenCode under the same conditions — a claimed +14.6% improvement attributed to the harness, memory, and verification gates rather than model differences. The project also documents readiness for Harbor evaluation and includes a Python adapter and trace output for headless benchmark runs.

Model Freedom and Extensibility

Metis is model-agnostic by design. Users can configure any supported LLM provider via API key or use the /login command for subscription providers. The TypeScript plugin system supports Agent Skills, MCP (Model Context Protocol) extensions, and custom packages. The npm package (@wholiver_hu/metis) exports a Node.js SDK and an RPC entry point for programmatic integration.

Update: v1.1.9

The latest release is v1.1.9, published on August 30, 2026. The repository was last pushed the same day, indicating active development. The project launched in July 2026 and has accumulated 104 stars and 12 forks in its first weeks. Desktop installers are available as a .dmg for macOS Apple Silicon and a .exe/.zip for Windows x64, bundling the CLI and server runtime without requiring a separate Node.js installation.