Skip to main content

Darwinian Harness

darwinian-harness is a local meta-harness for AI agent tools: one CLI to organize skills, MCP servers, extensions, defaults, project overlays, downstream tool configs, and diagnostics.

The CLI is drwn.

Agents are only as reliable as the harness around them. darwinian-harness makes that harness explicit, inspectable, reusable, and safe to write into downstream tools.

What it harnesses​

  • Skills and instructions that guide agent behavior
  • MCP servers and tool definitions that control capability access
  • Extensions such as Parallel, Beads, and MarkItDown that bundle project-level setup and diagnostics
  • Machine-wide defaults for reusable local capabilities
  • Project overlays for repository-specific agent behavior
  • Downstream state for Claude Code, Codex, Cursor, and ~/.agents
  • Diagnostics that report drift before mutating local files

Why this exists​

Local agent setups tend to drift. One tool gets a new MCP server, another has an older skill directory, and a project needs a slightly different harness than the global baseline.

The harness around an agent is usually scattered across dotfiles, skill directories, MCP configs, extension setup scripts, and project conventions. darwinian-harness gives those pieces a local control plane you can inspect, version, dry-run, and write deliberately.

It is useful when you want:

  • one reusable MCP and skill inventory instead of separately hand-edited tool configs
  • one harness layer shared across compatible agent tools
  • project-specific overrides without rewriting global config
  • diagnostics for stale links, drifted config, and missing generated files
  • an operator CLI that reports before it mutates

If you only need a single MCP config file for one tool, this project is probably more structure than you need.

What's next​