Introduction

HAR is an open-source CLI and MCP control plane for project-owned development harnesses. It makes a repository operable by coding agents without tying that repository to one model, editor, test framework, or hosted platform.

The problem

Launching a real application often depends on knowledge that is scattered across README files, package scripts, shell history, and one developer’s machine:

• which dependencies and infrastructure must start first;
• how ports, databases, and environment files are allocated;
• which checks define “done”;
• where logs, previews, traces, and screenshots are written;
• how to clean up without destroying someone else’s work.

Agents can attempt to rediscover those details on every task, but that is slow and risky. HAR captures them once in a versioned .har/ directory.

The boundary

HAR provides:

• editable launch, verification, inspection, reset, and teardown scripts;
• isolated agent slots backed by fresh Git worktrees;
• a machine-readable stage and artifact registry;
• consistent CLI and MCP operations over project-specific scripts;
• persisted run history and exact-tree validation records;
• optional commit gates, agent skills, and worktree guards;
• a local Mission Control dashboard.

HAR does not write application code, replace CI/CD, or require a particular test runner. Project-specific tools such as Playwright and RocketSim are optional stage templates that compile down to the same generic stage contract.

Who uses it

Humans can run har env ... from a terminal. MCP-capable agents call the equivalent har_* tools. Cursor, Claude Code, and Codex users can also install managed project workflows. All three paths execute the repository’s own .har/ scripts.

Open source and portability

The CLI, MCP server, harness runtime, and local Mission Control are open source. The .har/ directory remains in your repository and is editable even if you stop using HAR’s CLI. Hosted coordination belongs to HAR Cloud; it does not replace the portable local contract.