Skip to content
CMCodeMonkey / Harness Pro

01OPEN SOURCE / AGENT WORK

Make agent work observable.

CodeMonkey coordinates requests and reviews. Harness Pro defines and runs the agent. Together, they turn a task into work you can follow and inspect.

Explore the tools See the workflow

Already paired in a deployed CodeMonkey setup.

EXAMPLE RUN / 042

REVIEW READY

From request to evidence

  1. REQUEST

    Review feedback received

    CodeMonkey queues the task

  2. CONTRACT

    Agent scope compiled

    Harness Pro sets model, tools, limits

  3. RUN

    Worker completes

    Artifacts and trace are recorded

  4. REVIEW

    A person decides

    Evidence is ready to inspect

Every step has an owner.CM × HP

ONE WORKFLOW / TWO RESPONSIBILITIES

Coordinate the work. Define the run. Review the result.

02CODEMONKEY / ORCHESTRATION

A control room for every agent task.

CodeMonkey is a self-hosted task orchestrator. It brings requests into one place, launches dedicated Docker workers, and makes progress and results visible to the people responsible for delivery.

Explore CodeMonkey on GitHub
01

Bring work in

Receive direct tasks and repository review events, then keep each request in a visible queue.

02

Run the right worker

Launch a dedicated Docker worker with a repository checkout, role instructions, and configured credentials.

03

Keep the evidence

Follow runs in the dashboard, inspect artifacts and chat, and use the optional acceptance pipeline for build, test, SonarQube, and human review.

CODEMONKEY / SYSTEM MAP

From request to a reviewable result

CONCEPTUAL FLOW

01 / INTAKE

Work comes in

Manual tasks and repository review events enter one queue.

02 / TRUST BOUNDARY

Validate the request

Authenticate its source and check permissions.

03 / CONTROL PLANE

CodeMonkey routes the work

Track the task, choose the action, and dispatch an isolated worker.

04 / EXECUTION → 05 / OUTPUT

Agent worker→Run artifacts
Pipeline driver→Review gate
Chat worker→Dashboard

PERSISTENT STATE

Task records, pipeline state, and run history stay available.

This map shows the orchestration roles. The public repository setup guide lists the integrations and acceptance checks currently available.

03HARNESS PRO / EXECUTION

Define the agent before it runs.

Harness Pro is a provider-agnostic runtime for one headless agent run. Its versioned manifest spells out what the agent can access, which tools it can use, and when the run must stop.

Explore Harness Pro on GitHub
MANIFEST EXCERPT
{
  "provider": "anthropic",
  "model": "your-model",
  "capabilities": ["repo_read"],
  "turn_limit": 12
}

01 DECLARE

02 COMPILE

03 RUN + TRACE

Compilation validates the contract without accessing the workspace or calling a provider.

01

Declare the contract

A versioned JSON manifest chooses the provider, model, mounted paths, capabilities, MCP servers, behavior, and limits.

02

Compile before execution

Validate the manifest and inspect its runtime configuration without reading mounts or contacting a provider.

03

Run with a trace

Start one headless run with registered tools and record model and tool activity in a JSONL trace.

HARNESS PRO / SYSTEM MAP

From manifest to a traceable run

ONE AGENT / ONE RUN

  1. 01 / VERSIONED INPUT

    Agent manifest

    Select the model, mounted paths, capabilities, behavior, and run limits.

  2. 02 / SIDE-EFFECT-FREE

    Validate and compile

    Reject invalid fields and produce a deterministic runtime contract. No workspace access or provider call.

  3. 03 / RUNTIME CONTRACT

    Inspect the configuration

    See the compiled prompt, mounts, registered tool schemas, and limits.

  4. 04 / EXPLICIT LOCAL RUN

    Run one agent

    Resolve credentials, use the selected provider and registered tools, and enforce limits.

  5. 05 / REVIEWABLE OUTPUT

    Result and trace

    Read the final assistant text and a JSONL record of model and tool activity.

The manifest and compiled configuration define the run. The public repository documents the exact schema, capabilities, and supported providers.

04HOW THEY WORK TOGETHER

One task. Two systems. One reviewable result.

The handoff from orchestration to execution is explicit, and the outcome stays reviewable.

CODEMONKEY × HARNESS PRO

The handoff from task to decision

DEPLOYED FLOW

  1. 01 / CODEMONKEY

    Receive the request

    A manual task or repository event enters the queue.

  2. 02 / CODEMONKEY

    Dispatch the worker

    CodeMonkey passes task context and a checkout to the worker.

  3. 03 / HARNESS PRO

    Compile the contract

    The manifest defines the model, mounts, tools, and limits.

  4. 04 / HARNESS PRO

    Run the agent

    The runtime uses registered tools and emits a result and trace.

  5. 05 / CODEMONKEY

    Track the outcome

    Run status and worker artifacts remain visible.

  6. 06 / HUMAN REVIEW

    Decide what ships

    A person inspects the evidence and accepts or revises.

In use today: A deployed CodeMonkey setup already runs workers through Harness Pro. The public CodeMonkey checkout currently ships the Claude Code worker path; its Harness Pro adapter has not been published here.

05GET STARTED

Pick the part you need.

Run either project on its own. Their repositories have the full setup guides.

Start CodeMonkey

Configure a target repository, source-control access, and model credentials, then start the local dashboard.

TERMINAL
cp .env.example .env
./start.sh
Read the CodeMonkey setup guide

Start Harness Pro

Install the Python package, copy an example manifest, inspect the compiled contract, and run an agent.

TERMINAL
python -m pip install -e .
cp examples/standalone-anthropic.json my-agent.json
harness compile --manifest my-agent.json -o runtime.json
harness run --manifest my-agent.json --prompt "Review the repo" --trace-file trace.jsonl
Read the Harness Pro quick start