Skip to main content

​
Local Deployment

Run Myrm on your own machine with full control over your data.

​
One-liner install (fastest)

macOS / Linux / Git Bash
Windows (PowerShell)
Open http://localhost:3000 and add your LLM API key in Settings. Harness comes from PyPI; no separate harness repo required for self-host.

​
Clone and develop (OSS repo)

Dev scripts live under scripts/ in the open-source repo so contributors clone once and run locally.
Split dev tip: If you open only the WebUI (port 3000) without the backend, the UI shows a friendly banner with startup guidance—not cryptic errors. Settings cache offline and sync when the backend returns.

​
Prerequisites (Docker path)

  • Docker 24+ and Docker Compose v2
  • 4GB RAM minimum (8GB recommended)
  • 10GB disk space

​
Docker Compose

​
Configuration

Edit .env to set your API keys:
See Environment Variables for the full reference.

​
China Mainland Acceleration

The install script automatically detects China mainland networks (via timezone + connectivity check) and switches to domestic mirrors:
  • PyPI → pypi.tuna.tsinghua.edu.cn
  • npm → registry.npmmirror.com
  • Playwright → cdn.npmmirror.com
For Docker builds, pass --build-arg USE_CN_MIRROR=1:
Manual override: set MYRM_USE_CN_MIRROR=1 to force domestic mirrors, or MYRM_NO_CN_MIRROR=1 to disable auto-detection.

​
Verify Installation

Expected response:

​
Production Docker builds (OSS server image)

When building the open-source server image from pre-built harness wheels, the builder selects the correct platform core wheel via BuildKit TARGETPLATFORM and runs verify-harness-distribution before the image is considered ready. Mismatched or incomplete installs fail at build/startup time—not mid-conversation. See the harness distribution guide in the monorepo: myrm-agent-harness/harness_packaging/DISTRIBUTION_SYSTEM.md.

​
Modular engine packaging (embedders)

The harness ships as 24 lean runtime dependencies plus 11 optional Extras (browser, retrieval, ACP, observability, etc.). End users installing via install.sh, Docker, or Tauri get the full server bundle automatically—no manual extra selection. If you embed the engine in your own Python service, install only what you need:
Missing an Extra surfaces a clear install hint at call time (not a silent mid-conversation crash). Architecture tests lock the core list to prevent dependency creep—327 harness architecture tests verified (Jul 2026).

​
Automatic Dependency Management

Myrm handles dependencies automatically so you never need to manually install packages:
  • Channel SDKs — When you enable Matrix, Discord, or Feishu channels, the required Python packages are installed on-demand via an allowlist-secured mechanism. No restart needed.
  • MCP Servers — External tools run via npx/uvx which manage their own environments automatically. Just add a server in the GUI and it works.
  • Doctor diagnostics — Run myrm doctor to check system health: Python version, core dependencies, browser availability, and more. Each failure includes a fix suggestion.
  • Security by default — Only pre-approved packages can be auto-installed. MCP servers are scanned for malware (OSV) and prompt injection before activation.

​
Updating