agent-harnesses-mcp 0.1.0

MCP server for best-of-Agent-Harnesses: harness recommendations, search, and head-to-head decision guides over a hand-curated, weekly-rescored list of 110 agent harnesses

agent-harnesses MCP server

The best-of-Agent-Harnesses list as an MCP server, so agents can recommend harnesses instead of you reading 101 table rows.

Single file, stdio transport, no clone needed — it fetches harnesses.json from this repo at startup (or reads it locally from a checkout). Requires uv.

Install

Claude Code:

claude mcp add agent-harnesses -- uv run https://raw.githubusercontent.com/RyanAlberts/best-of-Agent-Harnesses/main/mcp/server.py

Any other MCP client (Cursor, Codex, Gemini CLI, ...):

{
  "mcpServers": {
    "agent-harnesses": {
      "command": "uv",
      "args": ["run", "https://raw.githubusercontent.com/RyanAlberts/best-of-Agent-Harnesses/main/mcp/server.py"]
    }
  }
}

Tools

Tool	What it does
pick_harness(use_case, max_complexity?, min_autonomy?, min_recovery?, open_source_only?, limit?)	Ranked recommendations for a use case, seeded by the list's hand-curated use-case index. max_complexity caps adoption surface (super simple → complex); min_autonomy requires a designed autonomy regime (step-gated → headless); min_recovery requires a failure-recovery tier (none → durable).
search_harnesses(query, limit?)	Keyword search across names, descriptions, tags, and categories.
get_harness(github_id)	Full record for one project.
list_comparisons()	The head-to-head decision guides (OpenClaw vs Hermes, terminal coding agents, …) with summaries.
get_comparison(slug)	Full markdown of one guide — architecture trade-offs, field reports, billing reality. Always current: served from the repo's main.
list_categories()	The 10 categories, use-case intents, and the complexity/autonomy/recovery scales.

Example: "pick_harness('sandboxed code execution for generated code', max_complexity='slightly complex', open_source_only=True)" → E2B, smolagents, Daytona... each with stars, tier, license signal, and a one-line reason.

Data is regenerated by scripts/generate.py; star counts carry a stars_captured date, and the comparisons index is rebuilt from comparisons/*.md on every refresh — the server always serves current main.

Distribution

The server is packaged as agent-harnesses-mcp (this directory's pyproject.toml) and registered in the official MCP registry as io.github.ryanalberts/agent-harnesses (server.json at the repo root). Once a release is published, package-manager installs work everywhere:

# any MCP client, via PyPI
uvx agent-harnesses-mcp
# Claude Code
claude mcp add agent-harnesses -- uvx agent-harnesses-mcp

Until then (and forever, as the zero-install path), the raw-URL one-liner at the top of this README works from any machine with uv.

Publishing (maintainer runbook)

Releases are automated by .github/workflows/publish-mcp.yml on a mcp-v* tag: it builds the wheel, publishes to PyPI via trusted publishing, and publishes server.json to the official MCP registry via GitHub OIDC.

One-time setup, then never again:

1. On pypi.org: create the project name agent-harnesses-mcp → Settings → Publishing → add a trusted publisher: owner RyanAlberts, repo best-of-Agent-Harnesses, workflow publish-mcp.yml. No API tokens.
2. Nothing for the MCP registry — GitHub OIDC from this repo authorizes the io.github.ryanalberts/* namespace automatically.

Per release: bump the version in mcp/pyproject.toml and server.json (the workflow fails loudly on mismatch), then git tag mcp-v<version> && git push origin mcp-v<version>.

Also indexed via smithery.yaml (submit the repo once at smithery.ai); other directories (Glama, PulseMCP, mcpservers.org) crawl the official registry.