Claude Code Harness plugin
Harness drives a task through plan, work and review: spec and Plans.md live in the repo and a separate reviewer checks the result.
- 3,207 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The plugin installs hooks and guardrail rules, writes plan files into the project and runs agent commands.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/claude-code-harnessDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The plugin installs hooks and guardrail rules, writes plan files into the project and runs agent commands.
This is third-party code. Review the repository files before installing.
What it does
The Harness plugin adds five workflow skills to Claude Code, Codex CLI, Cursor and Grok: `/harness-plan`, `/harness-work`, `/harness-review`, `/harness-sync` and `/harness-release`. The agent first drafts `spec.md` and `Plans.md` with completion criteria, then works through approved tasks while a separate reviewer checks the result. Risky operations (billing, external sends, secret reads, production deploys, force pushes) pass through guardrail rules written in Go. Larger jobs get `/breezing`, which splits work between agents, and `/harness-loop`, which repeats the cycle up to a limit of eight iterations.
Who it is for. Developers who hand whole tasks to an agent and want a plan, status and independent review instead of one long chat.
Good fit when
- A form or API change must be split into tasks with completion criteria before any code is written.
- The agent has finished and the result should go to a separate reviewer rather than the author.
- A session broke off midway and work resumes through `/harness-work --resume latest`.
- A team introduces agent rules: what needs permission, from reading secrets to production deploys.
Not a fit when
- A one-line fix where a plan and review only slow things down.
- You need full command isolation: Harness checks do not guarantee that every command is contained.
- The project cannot keep plans as files, so adding `spec.md` and `Plans.md` to the repo is unacceptable.
Example request
Plan the task "the form send button creates two messages on a double click" and show completion criteria before any work starts.Limitations
Claude Code needs version 2.1 or newer, and building from source needs Go 1.26. Guardrail checks cover operations that pass through hooks, and coverage varies by tool: Cursor has no filesystem isolation and native Codex child agents inherit parent permissions. A reviewer role alone is not read-only isolation. Before upgrading from the old `claude-code-harness` plugin, read the rename guide: the plugin ID is now `harness`. Setup scripts for Codex, Cursor and Grok run from a clone of the repository and replace existing `worker.toml` and `reviewer.toml`.
How to disable. Remove the plugin through `/plugin` in Claude Code and delete `spec.md`, `Plans.md`, the `.claude/state/` folder and files added by `/harness-setup` from the project. For Codex, Cursor and Grok follow the removal steps in the installation guide.
FAQ
How does `/harness-plan` differ from an ordinary agent plan?
`/harness-plan` writes `spec.md` and `Plans.md` with expected behavior, tasks, completion criteria and unknowns. You approve or correct them before work starts.
What does `/harness-review` do?
It requests a check from a reviewer separate from the implementation. Serious findings block completion while minor suggestions can remain.
Is Harness paid?
No, it is MIT licensed. You pay for your assistant subscription and model usage, which grows with more workers and retries.
How do I continue after a session breaks off?
Run `/harness-work --resume latest`. Ask the assistant to show what was recovered before relying on it.
Which assistants are supported?
Claude Code, Codex CLI, Cursor and Grok are listed as supported. OpenCode is partly compatible while Hermes Agent and GitHub Copilot CLI remain experimental.
Discussion
No questions or reviews yet.
Related
PluginsA DiDi plugin for Codex and Claude Code that runs a task from proposal to archive and keeps verified project knowledge in the repo
A Claude Code plugin with /readiness and /setup skills that scores a repo for agent work and installs rules, hooks and checks.
ADLC Harness takes a task from intent to PR with one command, and the human approves only four key gates.
A CLI, hooks and skills for long Claude, Codex and Pi work: the issue and PR own the task and review gets a PR with checks
Codely team skills that run agent work in phases: a plan in a file or GitHub issues, review after each phase and a clean commit
Morning Star drives agent work from request to merge through a PM agent, plan review and acceptance, with a decision log
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Claude Code Harness (this card) | no | 3,207 | |
| Superpowers | no | 316,679 | a skill set for methodical development; Harness adds spec and task files, a separate reviewer and guardrail rules |
| Compound Engineering | no | 26,054 | a loop that accumulates knowledge between tasks; Harness builds work around a plan, checks and review |
README in short
The README presents Harness as a plan, work, review loop built on `spec.md` and `Plans.md`. Installation goes through the Claude Code marketplace or setup scripts for Codex, Cursor and Grok.