Skip to content
Conduit

mcp server

harness-codex-plugin

Essential workflow framework runtime for Codex: custom agent harnesses with specs, artifacts, traces, and fail-closed verification.

EvanAI0331/harness-codex-plugincommunitylistedunreviewed
Use tool

Set up

Runs on your computer. Pick your agent and operating system.

This project does not publish an install method we can turn into a config. Follow the README instead.

README

synced Oct 29, 2026

Harness Codex Plugin

Official Codex Plugin Directory publishing is not yet self-serve. Harness is currently distributed through a GitHub-backed Codex marketplace.

codex plugin marketplace add https://github.com/EvanAI0331/harness-codex-marketplace.git --ref v0.1.8

Create and verify a workflow contract:

python3 scripts/harness_cli.py init-workflow --template software_delivery --output ./harness.workflow.json
python3 scripts/harness_cli.py verify-workflow ./harness.workflow.json

Harness is the workflow framework for custom Codex agent systems. It turns vague workflows into role-bound, execution-bound, output-bound harnesses. It prevents fake success, silent fallback, and uncontrolled script-only execution.

Harness Studio is the included visual workbench for building, inspecting, and running structured multi-agent harnesses. It pairs a single workspace with task instances, capability registries, artifact-based execution traces, and a task-oriented runtime so you can see what the system built, what it ran, and what it delivered.

This repository is an open-source OSS skeleton with a real runtime path, not a production-hardened platform and not a mock-only demo site.

Core Features

  • Codex Plugin packaging with skills, CLI checks, MCP tools, and repo marketplace distribution
  • Single workspace for harness generation and editing
  • Independent run instances with their own task instructions
  • Task instance planning that drives multi-agent execution
  • Capability registry and capability resolution
  • Artifact-based outputs for task instances, node outputs, tool results, and final deliverables
  • Runtime trace and event stream
  • Demo mode for offline or low-friction local exploration
  • SQLite-backed persistence

Why Harness

Codex needs a workflow layer when tasks grow beyond one-shot prompts. Harness supplies that layer: graph, specs, capabilities, run instances, artifacts, and trace. The result is a repeatable framework for custom agent work instead of freestyle execution.

Core Concepts

  • Harness Goal: The long-lived purpose of the harness. This belongs in the workspace intake area.
  • Run Task Instruction: The task for one specific run. This belongs in the run entry flow.
  • Build: Generates or rebuilds the harness graph, specs, and capabilities.
  • Run: Executes one task against a built harness.
  • Task Instance: The persisted run plan that agent execution reads from.
  • Artifact: The primary unit of progress and result storage.
  • Final Deliverable: The task result produced by the responsible agent.
  • Trace: The runtime event history for build and run.

Architecture

At a high level:

  • Workbench: src/components/HarnessWorkspace.tsx
  • Build orchestrator: src/lib/build-orchestrator.ts
  • Run orchestrator: src/lib/run-orchestrator.ts
  • Adapters: src/lib/llm, src/lib/specx, src/lib/scriptx, src/lib/capabilities, src/lib/runtime
  • Artifact layer: src/lib/artifact-repository.ts, src/lib/run-output/final-deliverable-aggregator.ts
  • Persistence: SQLite in src/lib/sqlite.ts
  • Streaming: SSE via src/lib/useEventStream.ts

Quick Start

1. Install

npm install

2. Configure

Copy .env.example to .env and adjust the values you need.

Important:

  • credentialRef is a server-side pointer to a secret environment variable.
  • Example: credentialRef=LLM_MAIN means the server reads LLM_MAIN_API_KEY.
  • The browser never receives raw API keys.

3. Demo Mode

npm run demo

Demo mode keeps the orchestrator real and swaps only adapter implementations so you can explore the flow without live credentials.

4. Run the App

npm run dev

Then open:

  • /harness/[id] for the workspace
  • /runs/[runId] for run results
  • /harness/settings for runtime settings

5. Smoke Test

npm run test:smoke

The smoke test runs the app in DEMO_MODE=true, verifies SQLite initialization, harness creation, build, run, artifact retrieval, and final deliverable availability.

Demo Scenario

The repository includes a public demo flow built around a Repository Audit Harness.

Typical flow:

  1. Create a harness
  2. Enter a Harness Goal
  3. Generate Harness
  4. Run New Task
  5. Enter a Run Task Instruction
  6. Open /runs/[runId]
  7. Inspect Final Deliverable, Final Report, Artifacts, and Runtime Trace

Pages

  • /harness/[id] - main workspace
  • /runs/[runId] - run detail view
  • /harness/settings - runtime settings

Current Limitations

  • This is an OSS skeleton, not a production-hardened multi-tenant service.
  • Default adapters may use demo/mock implementations when DEMO_MODE=true.
  • Some integrations are pluggable placeholders and may need real credentials or local setup.
  • Security boundaries, tenancy isolation, and hardening are still in progress.

Scripts

  • npm run dev
  • npm run demo
  • npm run build
  • npm run start
  • npm run lint
  • npm run typecheck
  • npm run plugin:verify
  • npm run plugin:verify-specs
  • npm run test:smoke
  • npm run db:reset

Codex Plugin

The plugin entrypoint is .codex-plugin/plugin.json.

Codex skills:

  • harness-workflow-builder
  • harness-runtime-verifier
  • harness-plugin-packager

MCP tools:

  • harness.verify_plugin
  • harness.verify_specs
  • harness.explain

All plugin verification commands fail closed. Missing specs, plugin metadata, skills, marketplace metadata, or MCP wiring return ok=false with failure_state and details.

Documentation

Third-Party Notices

Vendored dependencies are documented in THIRD_PARTY_NOTICES.md.

↑↓ navigate openesc close