agent-harnesses MCP server

mcp-name: io.github.RyanAlberts/agent-harnesses

The best-of-Agent-Harnesses list as an MCP server, so agents can recommend harnesses instead of you reading 100+ 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

Published on PyPI and the official MCP registry as io.github.RyanAlberts/agent-harnesses. Claude Code:

claude mcp add agent-harnesses -- uvx agent-harnesses-mcp

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

{
  "mcpServers": {
    "agent-harnesses": {
      "command": "uvx",
      "args": ["agent-harnesses-mcp"]
    }
  }
}

No-install alternative — run the single source file straight from this repo:

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

Tools

Example: "recommend('an always-on personal assistant that lives in my chat apps', open_source_only=True)" → one top pick with the reason, two alternatives, anything to avoid for this need, and the guide to read next.

Example: "compare(['openclaw/openclaw', 'NousResearch/hermes-agent'])" → both records side by side, who leads on which axis, and a pointer to the OpenClaw vs Hermes guide.

Example: "compare_for('an always-on personal assistant in my chat apps')" → OpenClaw vs Hermes vs Khoj with each pick's ranking reason and the axis edges — OpenClaw leads stars, Hermes is simplest to adopt, Khoj takes the sandboxing edge (its generated code always runs isolated; the others' sandboxes are opt-in), OpenClaw and Hermes share the lifecycle-hooks edge — plus the OpenClaw vs Hermes decision guide to read next.

Example: "compare_for('sandboxed code execution for generated code')" → E2B vs Daytona vs smolagents side by side with why_picked reasons for each.

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 live in the official MCP registry as io.github.RyanAlberts/agent-harnesses (server.json at the repo root), which directories like Glama and PulseMCP crawl. The registry validates PyPI ownership via the mcp-name: marker at the top of this README — keep it.

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>.

Directories like Glama, PulseMCP, and mcpservers.org crawl the official registry — no per-directory submissions needed. (Smithery's current publish flow takes hosted-HTTP servers or .mcpb bundles, not stdio-from-GitHub, so this server isn't listed there by design.)