Agent Harness Kit: structured agent workflow
Agent Harness Kit
The ahk CLI sets up five agent roles, a task MCP server and spec skills in your project for Claude Code, Codex and Cursor
- 184 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Initialization creates agent files, skills and an MCP entry in the project, and the health script runs project commands
- Roles barred from editing are limited by the agent's tools and not by a sandbox, so there is no isolation from dangerous commands
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/agent-harness-kitDetects 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.
- Initialization creates agent files, skills and an MCP entry in the project, and the health script runs project commands
- Roles barred from editing are limited by the agent's tools and not by a sandbox, so there is no isolation from dangerous commands
This is third-party code. Review the repository files before installing.
What it does
Agent Harness Kit, or AHK, gives different AI agents a shared process: documents, tracked tasks, handoffs and health evidence. `ahk init` in a project directory asks for the provider, docs path and storage mode, then creates configuration, role files and skills for the chosen agent: Claude Code, Codex CLI, Cursor, Grok Build, OpenCode or OpenCode V2. Five roles handle a task. The lead claims it, the explorer studies the code, the consultant compares approaches, the builder writes, and the reviewer checks and sends work back to the builder with corrections. Task claims, the action log, handoffs and the final project health check live in an MCP server, while documents stay ordinary files. Individual skills start on explicit request: `ahk-use-case` and `ahk-spec` prepare a scenario and a functional spec, `ahk-spec-tech` designs the implementation, `ahk-triage` finds a bug's cause without edits, `ahk-test` writes tests only, `ahk-docs` explains AHK itself.
Who it is for. Teams and solo developers who run tasks through Claude Code or Codex and want a trail: who took a task, what was verified, what was approved.
Good fit when
- A feature is vague and a scenario and spec approved by a human should come before code
- A bug has to be analyzed from evidence without touching code, and a report produced
- Several agents serve one repository and need shared task tracking
- A finished piece of code goes through review by a separate role instead of the agent that wrote it
- Before a task closes, the server must run the project health script and keep the result
Not a fit when
- A small edit in one file: roles and documents only slow the work down
- There is nothing to build a health script from: a stub that always succeeds does not count as completion evidence
- You need isolation from dangerous commands: roles restrict tools but are not a sandbox
- Tasks arrive from Jira: AHK does not fetch cards automatically and the request text is pasted in by hand
Example request
Use ahk-feature to turn the Jira card pasted below into a functional spec. Save a draft, ask for approval and do not create a task.Limitations
The starter health script that `ahk init` creates must be replaced with commands that actually verify the project. Role permissions are set by the agent's tools: prompt instructions, Codex ones included, are not a security boundary.
Yarn Berry has no `yarn global`, so global installation goes through pnpm, npm or Bun. After initialization the provider is restarted so it reloads its MCP configuration.
OpenCode V2 support was exercised on version 2.1.4 with a deterministic local model provider, which does not establish quality with paid models. The repository's last commit is from October 2026.
How to disable. Remove the package with `npm uninstall -g @cardor/agent-harness-kit`, and delete the role files, skills and MCP entry from the project: for Claude Code that is `.mcp.json` and the agents directory.
FAQ
How do I install AHK and start?
Run `pnpm add -g @cardor/agent-harness-kit`, then `ahk init` in the project directory. The wizard asks for the provider and docs path. After that restart the agent and make a first bounded request, for example through `ahk-feature`.
Which agents are supported?
Claude Code, Codex CLI, Cursor, Grok Build, OpenCode and OpenCode V2. For each, `ahk init` creates its own files: for Claude Code that is `.mcp.json` in the project root.
How does ahk-triage differ from ahk-ask?
`ahk-triage` diagnoses a bug and returns an evidence-backed report without changing code. `ahk-ask` answers questions about project code and does not touch task state. Both are read-only.
Which agent roles does AHK have?
Five: the lead coordinates, the explorer studies code, the consultant compares approaches, the builder implements, and the reviewer checks and returns corrections to the builder. The final project health check runs on the server.
Where does AHK keep state?
Tasks, the action log, handoffs and the health result are kept by the MCP server, and documents stay as files in the chosen directory. The database can live globally in the home storage or locally, commonly in `.harness/harness.db`.
Discussion
No questions or reviews yet.
Related
CLI toolsGives each coding agent its own repository copy with separate ports and a database, then verifies results in one shared pipeline.
Over twenty skills for coding agents: MCP and hooks, plan mode, a bug-fix protocol, prompting and a bash safety hook
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
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
Repository Harness installs an AGENTS.md entrypoint, a document map and explicitly invoked onboarding skills into a repo and updates them safely
A terminal coding agent: plan mode before edits, settings import from Claude Code and Codex, skills, MCP, any model provider
README in short
The README gives a table of what you need, what to ask for, what you get and what it does not do for each skill, and two process diagrams: idea to spec and task to verified work. Installation, storage, per-role model choice and OpenCode V2 specifics follow.