Open-source agent infrastructure

Give coding agents a real place to work.

HAR turns any repository into a reproducible workspace with isolated worktrees, project-owned commands, and evidence that changes actually pass.

Works with
Claude CodeCursorCodexAny MCP agent
agent-session / slot-03 running
A
Agentrequests task
Repositoryreads .har/
01
Create slotisolated worktree
02
Launch stackports + services
03
Make changesinside worktree
04
Verifyproject checks
Validated branchlogs · artifacts · tree hash
ready
01
Read the project contractAgent discovers the repository through .har/
02
Create an isolated slotFresh worktree, ports, services, and environment
03
Launch, change, verifyRun the same workflow your team trusts
Hand off evidenceLogs, artifacts, exact validated tree hash
agentSlots4
last runpassed in 31.2s
01

Project-owned

The workflow lives with the code, not inside a vendor dashboard.

02

Agent-agnostic

One stable contract for CLI users, MCP agents, and future tools.

03

Evidence-first

Every run can leave logs, artifacts, status, and a validated tree hash.

The missing operating layer

Your agent should not have to rediscover how your repository works.

Most coding agents are good at editing files. The hard part is everything around the edit: launching the stack, isolating parallel work, running the right checks, and producing a trustworthy handoff.

01 / Contract

The repository explains itself.

HAR scaffolds a human-readable, machine-readable .har/ folder. Your team owns it, reviews it, and evolves it with the project.

Explore the harness contract
.har/project-owned
my-app/
├── AGENT.md
└── .har/
    ├── README.md
    ├── stages.json
    ├── launch.sh
    ├── verify.sh
    ├── harness.env
    ├── logs/
    └── artifacts/
editable like any other code
02 / Isolation

Parallel work without collisions.

Each agent gets a dedicated worktree and environment slot.

main
slot-01agent A
slot-02agent B
slot-03agent C
03 / Verification

Proof, not “looks good.”

Normalized stages surface exactly what ran, how long it took, and what passed.

factory-lint5.3spassed
unit-tests25.7spassed
browser-e2e41.2spassed
One workflow, every agent

From “take this task” to a validated branch.

HAR normalizes the infrastructure around coding so the agent can focus on the code—and the reviewer can trust the result.

01 — Discover

The agent reads one stable interface.

Instead of guessing project-specific shell commands, the agent asks HAR what the repository can launch, verify, inspect, reset, and tear down.

$har env status --json
HAR MCPproject description
toolhar_describe_projectrepo: /workspace/my-app
{
  "profile": "default",
  "agentSlots": 4,
  "stages": [
    "launch", "verify", "test"
  ]
}
Mission Control

See the factory, not just the final diff.

A local dashboard for repositories, active slots, run history, verification stages, logs, and artifacts. The open-source control plane stays close to your machine and your code.

  • Observe every agent slot in one place
  • Inspect deterministic verification stages
  • Keep local and self-hosted workflows portable
Read about Mission Control
localhost:3847 / mission-controlHAR
HAR Mission Control repository dashboard
Active slots27across 6 repositories
Latest verification All checks passed31.2 seconds ago
A deliberately narrow layer

HAR coordinates the work around the model.

HAR is
  • A project-owned runtime contract
  • An agent-agnostic CLI + MCP surface
  • A reproducible launch and verification layer
  • A record of runs, artifacts, and evidence
HAR is not
  • A coding model
  • A replacement for CI/CD
  • A new test runner
  • A hosted platform lock-in
From the HAR journal

Building the operating layer for agentic software.

View all writing
agent
worktree
verify

Why coding agents need a harness, not another prompt.

The gap between generating code and safely operating inside a real repository.

.har/

The repository should own the agent contract.

Why launch, verification, and teardown belong beside the source code.

03

Evidence-first development for autonomous agents.

Move from confidence theater to deterministic, reviewable handoffs.

Start with one repository

Make your project agent-ready in minutes.

Install HAR, scaffold the contract, and let your coding agent work inside a reproducible environment.

Read the quick start