Harness To
FreeNot checkedExposes internal tools from agent harnesses (Claude Code, Codex, etc.) as a standard MCP server by intercepting LLM API calls.
About
Exposes internal tools from agent harnesses (Claude Code, Codex, etc.) as a standard MCP server by intercepting LLM API calls.
README
Contents: Features | Install | Demo | How it works | Notes
Turn any agent harness (Claude Code, Codex, OpenClaw, OpenCode, ...) into an MCP server.
How: harness_to_mcp sits between the harness and its LLM API, grabs the tool list from the hijacked request, and routes MCP tools/call back into the harness tool loop.
▮ Features
- ☑ One command to expose
claude/codex/openclaw/opencodeas an MCP server - ☑ Co-locates one MCP HTTP server and one hijack LLM API server on the same port
- ☑ Extracts the harness tool list automatically from intercepted LLM requests
- ☑ Mirrors captured harness system prompt into MCP
initialize.result.instructions - ☑ Forwards MCP
tools/callinto the harness tool loop and maps the tool result back - ☑ Compatible LLM API protocols:
- OpenAI Chat Completions (
openclaw,opencode) - OpenAI Responses (
codex) - Anthropic Messages (
claude)
- OpenAI Chat Completions (
- ☑ Isolated harness config — will not pollute the user's own config and logs
- ☑ One harness process per MCP session, the harness process is stopped automatically on session close
- ☑ Plain server mode for bringing your own harness
- ☑ Pure-Python, clean dependencies — easy to hack and vibe-code on
▮ Install
pip install harness_to_mcp
The target harness CLI (claude, codex, openclaw, ...) needs to be installed separately and available on PATH.
▮ Demo
1. One-liner to expose a harness as MCP
harness_to_mcp claude
# or: harness_to_mcp codex / openclaw / opencode
Each command starts a server and launches one corresponding harness instance. The harness is started with an isolated config, so it will not touch the user's own config or logs.
The MCP endpoint is then ready at:
http://127.0.0.1:<port>/mcp
Point any MCP client (Claude Desktop, Cursor, your own script, ...) at it and the harness's internal tools show up as standard MCP tools.
You can inspect the exposed tools with python examples/list_tools.py.
2. Only run the server (plug in any harness)
harness_to_mcp
This mode starts only the server. It listens on MCP plus all hijack LLM API routes, but does not launch any harness by itself. Configure your harness's LLM API as the hijack API, send one request, and its internal tools are exposed on MCP. This is also how you plug in claude / codex / openclaw / opencode with your own custom config.
Exposed endpoints:
| Purpose | Path |
|---|---|
| MCP | POST /mcp (alias: POST /harness_to_mcp/mcp) |
| OpenAI Chat Completions | POST /harness_to_mcp/v1/chat/completions |
| OpenAI Responses | POST / WebSocket /harness_to_mcp/v1/responses |
| Anthropic Messages | POST /harness_to_mcp/v1/messages |
3. Python API
We also provide a Python interface:
from harness_to_mcp import HarnessToMcp
with HarnessToMcp(port=9330) as server:
print(server.mcp_url) # e.g. http://127.0.0.1:9330/mcp
print(server.hijack_base_url) # e.g. http://127.0.0.1:9330/harness_to_mcp
▮ How it works
sequenceDiagram
participant C as MCP Client
participant M as harness_to_mcp<br/>MCP Server
participant H as harness_to_mcp<br/>Hijack API Server
participant R as Harness
C->>M: initialize
M->>R: launch harness session
R->>H: send hijacked LLM API request
H-->>M: extract tool list from request
M-->>C: tools/list
C->>M: tools/call
M-->>H: resolve waiting request as tool call
H-->>R: return tool call payload
R->>R: execute internal tool
R->>H: send next request with tool result
H-->>M: match tool_call_id and deliver tool result
M-->>C: MCP tool response
In short:
- MCP client calls
initialize→ we spawn a harness. - The harness fires off its first LLM request (with the tool schema in it) → we intercept the request, extract the tool list, and reply to the MCP client with
tools/list. - MCP client calls
tools/call→ we complete the pending LLM response as a tool call, the harness executes its internal tool, and sends the result back in the next LLM request. - We match the
tool_call_id, extract the tool result, and return it as the MCP tool response.
▮ Notes
- The LLM API layer is split into reusable adapters:
chat completions,responses,messages. - The harness launcher layer is split per-harness:
opencode,openclaw,codex,claude. - Plain server mode (
harness_to_mcpwith no subcommand) never auto-launches a harness. - Intercepted waiting requests are kept alive with periodic heartbeat bytes until MCP receives the next tool call.
▮ License
Install Harness To in Claude Desktop, Claude Code & Cursor
unyly install harness-to-mcpInstalls into Claude Desktop, Claude Code, Cursor & VS Code — handles npx, uvx and build-from-source repos for you.
First time? Get the CLI: curl -fsSL https://unyly.org/install | sh
Or configure manually
Run in your terminal:
claude mcp add harness-to-mcp -- uvx harness_to_mcpFAQ
Is Harness To MCP free?
Yes, Harness To MCP is free — one-click install via Unyly at no cost.
Does Harness To need an API key?
No, Harness To runs without API keys or environment variables.
Is Harness To hosted or self-hosted?
Self-hosted: the server runs locally on your machine via the install command above.
How do I install Harness To in Claude Desktop, Claude Code or Cursor?
Open Harness To on unyly.org, pick your client tab (Claude Desktop, Claude Code, Cursor) and press Install — the config is generated automatically, no JSON editing.
Related MCPs
GitHub
PRs, issues, code search, CI status
by GitHubFilesystem
Secure file operations with configurable access controls.
Memory
Knowledge graph-based persistent memory system.
Template MCP Server
A CLI tool to create a new Model Context Protocol server project with TypeScript support, dual transport options, and an extensible structure
by mcpdotdirectCompare Harness To with
Not sure what to pick?
Find your stack in 60 seconds
Author?
Embed badge for your README
Browse similar
All development MCPs
