agent-harnesses — Claude Code MCP server: pick an agent harness
An MCP server that tells your agent which harness fits your model and task. Useful when you are choosing a runtime for an agentic system instead of guessing from star counts.
claude mcp add agent-harnesses -- uvx agent-harnesses-mcpWhat agent-harnesses does
The server hands your agent a curated catalog of agent harnesses, orchestration frameworks and harness techniques. Two tools do the work: recommend and pick_harness. Your agent calls them and gets a harness matched to the model and the job at hand, rather than inheriting whatever someone else benchmarked. The catalog also carries decision guides, setup templates and skills that audit your own agent setup.
It runs as a stdio server through uvx, so uvx has to be present on the machine. Claude takes a single command; Cursor and GitHub Copilot need the JSON block pasted into their config. No API key is involved.
The catalog's argument is blunt: harness quality moves results further than a model swap, and rankings barely carry over between models. So the harness is a pairing with your model, and it has to be re-asked whenever the model changes. If your stack is settled and the model is not going anywhere, this server adds little.
What it helps with
- match a harness to a specific model and task
- compare orchestration frameworks by autonomy and recovery
- find a setup template for an agent runtime
- audit what your agent can reach and what it spends
- track newly added harnesses and weekly star gains
Example prompts for the agent
- Pick a harness for Opus on a large-repo refactor
- Call
pick_harnessand explain why that one beats the alternatives - Audit my current agent setup: what it can reach and what it spends
How to install agent-harnesses
Platform CLI
claude mcp add agent-harnesses -- uvx agent-harnesses-mcpuvx runs the command named after the package; if the README names another one, use uvx --from agent-harnesses-mcp <command>
Platform CLI
codex mcp add agent-harnesses -- uvx agent-harnesses-mcpuvx runs the command named after the package; if the README names another one, use uvx --from agent-harnesses-mcp <command>
Config file
{
"mcpServers": {
"agent-harnesses": {
"command": "uvx",
"args": [
"agent-harnesses-mcp"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from agent-harnesses-mcp <command>
Config file
{
"servers": {
"agent-harnesses": {
"type": "stdio",
"command": "uvx",
"args": [
"agent-harnesses-mcp"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from agent-harnesses-mcp <command>
Platform CLI
gemini mcp add agent-harnesses uvx agent-harnesses-mcpuvx runs the command named after the package; if the README names another one, use uvx --from agent-harnesses-mcp <command>
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about agent-harnesses
How do I install agent-harnesses in Claude Code?
Run claude mcp add agent-harnesses -- uvx agent-harnesses-mcp. The server then shows up in your Claude MCP server list.
What does agent-harnesses do?
It gives your agent a catalog of agent harnesses plus the recommend and pick_harness tools, so the agent can choose a runtime for your model and task.
What do I need before it runs?
uvx on the machine, since the server starts as uvx agent-harnesses-mcp. No API keys are required.
