Metis AI 发布自托管智能体工作台，支持多模型供应商与 MCP 网关

Metis AI is designed for people who want an agent workspace they can run, configure and extend themselves — without tying the entire experience to a single model provider.

Most AI interfaces stop at a chat transcript. Metis AI treats a conversation as a durable workspace. It combines the agent, its context, the tools it can use and the artifacts it creates in one private, self-hosted environment.

Feature overview

Metis AI includes the following capabilities:

The Windows desktop Remote Client is maintained in the public metis-remote-client repository. Its installer and automatic updates come from that repository's GitHub Releases. Existing clients that still check their paired Metis server receive those public release files through an authenticated proxy.

Tutorial & introduction

The video walks through installation, provider setup, workspaces, notes, automations, browser control, voice input, sharing, agent modes, MCP servers, remote clients and memories:

Watch the tutorial and introduction on YouTube

Versionierte Docker-Releases

Für produktive Installationen ist der versionierte Docker-Installer der empfohlene Pfad. Er lädt ein unveränderliches Image aus GHCR, erstellt keine Git-Checkout- Abhängigkeit und legt Daten sowie Workspace außerhalb des Containers ab.

curl -fsSL https://github.com/f1shyondrugs/metis-ai/releases/download/v1.0.10/metis-docker-install.sh \\
  -o metis-docker-install.sh
bash metis-docker-install.sh --version v1.0.10

Für reproduzierbare Installationen kann ein konkreter Release-Tag verwendet werden:

bash metis-docker-install.sh --version v1.0.8

latest wird vor dem ersten Schreibzugriff in den zuletzt veröffentlichten stabilen Release-Tag aufgelöst; danach bleiben Image, Compose-Konfiguration und Manifest auf genau diesem Tag. Für reproduzierbare Installationen kann der Tag direkt mit --version vX.Y.Z gesetzt werden.

Ein Upgrade wird mit demselben Befehl und einer neuen Version ausgeführt. Die Installationsdatei .metis-release.json, .env, das Datenverzeichnis und der Workspace bleiben dabei erhalten; der Installer ersetzt nur den Image-Tag und startet die Compose-Services mit docker compose pull und up -d neu.

One-line installer

The versioned Docker release installer above is recommended for production. This legacy cross-platform bootstrap remains useful for development checkouts or native installations. On Linux it installs native systemd; pass -- --docker for Compose. macOS and Windows also default to native background hosts; use --docker / -Docker to opt into Docker Compose. The one-liner downloads the platform installer to a temp file and executes that file — it does not run the installer from a pipe.

/bin/bash -c "$(curl -fsSL https://github.com/f1shyondrugs/metis-ai/releases/latest/download/metis-install.sh)"

If an older native Linux install reports update status 127, rerun the current installer once with that machine's existing installation directory (and its service name if customized). The old, already-running updater cannot acquire this PATH repair before its first successful update:

/bin/bash -c "$(curl -fsSL https://github.com/f1shyondrugs/metis-ai/releases/latest/download/metis-install.sh)" -- \
  --non-interactive --native --install-dir "/absolute/path/to/metis-ai"

This keeps .env and the data directory but replaces tracked local checkout changes with the selected upstream version. Native installs now save the Node and pnpm directories in .env for service starts and future updates.

Uninstall with the same bootstrap:

/bin/bash -c "$(curl -fsSL https://github.com/f1shyondrugs/metis-ai/releases/latest/download/metis-install.sh)" -- uninstall --yes --keep-data

The same command works on macOS and Linux. Do not use curl | bash against linux.sh or macos.sh directly.

The default path never creates an account or asks for credentials. Start the app and create the first account in the first-run UI. For agents and CI, pass optional configuration flags after --:

/bin/bash -c "$(curl -fsSL https://github.com/f1shyondrugs/metis-ai/releases/latest/download/metis-install.sh)" -- \
 --non-interactive --port 3100

Use --help for the available options.

On Windows, the one-liner is a bootstrap without a param() block so Invoke-Expression is valid. It saves windows.ps1 and invokes it with -File:

irm https://github.com/f1shyondrugs/metis-ai/releases/latest/download/metis-install.ps1 | iex

For a prompt-free Windows installation, download the platform script and pass named arguments:

irm https://raw.githubusercontent.com/f1shyondrugs/metis/master/install/windows.ps1 -OutFile install.ps1
.\install.ps1 -NonInteractive

During a normal interactive installation, every installer asks for the installation directory before it creates files or installs dependencies. Press Enter to accept ~/metis-ai, or enter a different path.

All three installers also accept argument-only configuration for the install directory, data directory, agent workspace, ports, bind address, service name and public URL. Use --help on Linux/macOS or -Help on Windows for the complete list.

Windows registers Metis AI in the Start menu and Installed Apps. A windowless host starts at sign-in, supervises the web app, worker and MCP gateway, and writes logs to the data directory. The Start menu entry opens the interface in your default browser. macOS installs a Finder/Spotlight app under ~/Applications; launchd runs the native services in the background. Create your account in the browser; installer password arguments remain optional for automation.

Both desktop installers check the effective web/MCP ports before installing dependencies or building, and again after restoring an existing configuration. A conflict reports the port and owning process; use -Port / -McpPort on Windows or --port / --mcp-port on macOS to choose free ports. Existing configuration is preserved on upgrade, so change its .env ports when needed. An unrelated HTTP 200 response is not accepted as Metis health.

Non-interactive options

The option names are intentionally listed side by side so the same deployment configuration can be reproduced on every supported operating system. Linux and macOS use Bash options; Windows uses PowerShell named parameters.

For Linux and macOS, pass installer arguments after -- to /bin/bash -c "$(curl ...)". For PowerShell, download install/windows.ps1 and invoke it with -File. -SkipRuntimeInstall only skips Windows' automatic Git/Node.js installation; it still verifies that the required tools are available.

The installer uses safe defaults for machine-specific values and does not create the first user; read the operator token from DATA_DIR/setup-token locally and enter it in the first-run UI. The app creates this private file before listening. Review downloaded scripts before executing them in security-sensitive environments. The repository source can be overridden with METIS_AI_REPO_URL.

The installers also ask whether the web application should be reachable on the local network. The secure default binds to 127.0.0.1; choosing the network option binds to 0.0.0.0 and requires a firewall or trusted TLS reverse proxy.

The scripts are hosted in the repository under install/; the website and Nginx configuration are not required for installation. Every installer writes an installation manifest and a matching uninstaller into the selected installation directory. Use --keep-data (or -KeepData on Windows) to remove services and application files while retaining chats, memories, uploads and encrypted provider credentials. Use --dry-run before removal when reviewing an existing installation.

• Node.js 22+
• pnpm 9+
• A supported AI provider credential, unless you only want to explore the UI

1. Install

git clone https://github.com/f1shyondrugs/metis.git
cd metis-ai
pnpm install

2. Configure

cp .env.example .env

Edit .env and set at least a login username/password. For provider credentials, configure connections later from Settings → Providers. Never commit .env.

3. Run the development server

pnpm dev

Open http://127.0.0.1:3100 and sign in with the credentials from .env.

Docker Compose

With Docker and Compose installed, start the app, worker and MCP gateway with:

docker compose up --build

Published ports follow AI_CHAT_HOST or AI_CHAT_BIND in .env (default 127.0.0.1). Set either to 0.0.0.0 for LAN access, then apply with ./reload.sh in the install directory. docker compose restart keeps the old bind address and environment.

When METIS_WORKSPACE is unset, Compose mounts ./workspace at /workspace. Set METIS_WORKSPACE to use a different host workspace directory.

Production-style start

pnpm build
pnpm start

The custom server listens on AI_CHAT_HOST (default 127.0.0.1) and defaults to port 3100. Set AI_CHAT_HOST=0.0.0.0 only when the app should be reachable on the local network, and put it behind an authenticated TLS reverse proxy before exposing it beyond a trusted LAN.

Provider connections

Provider connections are managed from Settings → Providers. Supported connection types include:

• Cursor
• OpenAI, Anthropic, Google Gemini and xAI/Grok
• OpenRouter
• Ollama and other local endpoints
• Codex, Claude Code and supported Antigravity credentials
• Generic OpenAI-compatible APIs such as Groq, DeepSeek, Mistral, Together, vLLM, LM Studio and LiteLLM

API keys and supported account credential bundles are encrypted at rest and are not returned to the browser. Set AI_CHAT_SECRETS_KEY before saving a connection:

openssl rand -hex 32

The key must represent exactly 32 bytes: 64 hexadecimal characters or a 32-byte base64 value. Google Vertex/ADC connections additionally need a GCP project and configured Application Default Credentials. The optional Antigravity Python path needs:

python3 -m pip install google-antigravity

The gateway lives behind the public module boundary in packages/mcp-gateway. It connects the agent runtime to registered local or remote MCP servers and supports discovery, workflows, web/documentation tools and selected platform integrations.

For a trusted deployment:

1. Set a long, random MCP_BEARER_TOKEN.
2. Keep the gateway on localhost or place it behind a trusted authenticated proxy.
3. Set AI_CHAT_ROOT, AI_CHAT_MCP_STATE_DIR and AGENT_CWD explicitly.
4. Enable remote, optional or dangerous integrations only when you understand their permissions.
5. Treat shell, filesystem, Docker and service-control tools as privileged.

See SECURITY.md, docs/PRODUCTION-AUDIT.md, and the gateway README.md before exposing any endpoint.

The complete example is in .env.example. The most useful settings are:

Production installs should run the static-only Scrapling service from deploy/systemd/metis-scraper.service.template. SearXNG is optional: search uses SearXNG first, then the local Scrapling public-search parser, then the configured Exa child server. Page fetches use Scrapling first and Exa only for failed or incomplete public-page extraction. Interactive, authenticated, or challenge pages belong in the persistent browser instead.

pnpm dev                 # Start Next.js in development mode
pnpm typecheck           # TypeScript validation
pnpm run test:providers  # Provider adapter tests
pnpm test                # Provider and security tests
pnpm lint                # ESLint
pnpm build               # Production build

Before opening a pull request, run the checks that cover your change. The project's contribution expectations are documented in CONTRIBUTING.md.

app/                  Next.js routes, pages and API handlers
components/           React UI and workspace panels
docs/                 Project documentation (start at docs/README.md)
lib/                  Agent runtime, providers, storage and MCP internals
packages/mcp-gateway/ Public MCP gateway module boundary
scripts/              Security and maintenance checks
tests/                Provider-focused tests
public/               Static assets and prompt data

Operator and architecture notes, including the current production audit, live in docs/.

Metis AI can execute powerful operations through agents and MCP servers. Do not expose a default installation directly to the public internet. Use strong secrets, a trusted proxy, least-privilege MCP configuration and isolated working directories. See SECURITY.md for the reporting process and deployment guidance.

Metis AI is released under the MIT License.