Skip to content

Bring your own coding agent

harnessforge generates files that every major coding agent already knows how to read. Zero per-IDE config required.

Claude Code

Reads .claude/CLAUDE.md automatically on session start. No setup needed.

Cursor

Reads .cursor/rules automatically. If you also want harness verify available as a tool, connect the MCP server (see below).

Codex CLI (OpenAI)

Reads AGENTS.md at the repo root. Standard.

Continue

Reads .continue/config.json. The harness adapter sets systemMessage + the project's recommended MCP servers.

Windsurf

Reads .windsurf/rules. Same shape as Cursor — sections for "always follow", "never touch", "never run", etc.

Gemini CLI

Reads AGENTS.md. Same as Codex CLI.

Aider

Reads AGENTS.md and the project's CONTRIBUTING.md. Works out of the box.

OpenHarness

Discovers MEMORY.md, CLAUDE.md, and SKILLS/ automatically. Run oh from the repo root and it picks them up.

Hermes / OpenClaw

Both read SOUL.md and the SKILLS/ directory (anthropics/skills format). Drop the agent in the repo and it picks up the right tone + procedures.

Connect harness mcp

For typed tool access from any MCP client:

// ~/.config/claude-desktop/config.json or ~/.cursor/mcp.json
{
  "mcpServers": {
    "harness": {
      "command": "uvx",
      "args": ["--from", "harnessforge[mcp]", "harness", "mcp"]
    }
  }
}

The MCP server exposes:

  • harness_inspect — full InspectionReport JSON
  • harness_blueprint_list — installed blueprints
  • harness_skills_list — local + catalog skills
  • harness_verify — run validators (same JSON as harness verify --json)
  • harness_profile_read — parsed .harness/profile.yaml

Your coding agent can call any of these as a typed tool.