Get Started

Overview: The Harness Runtime & Open Company

Learn how Munk AI coordinates your agent team across Code and Projects with hard quality gates, human checkpoints, and the open-source Open Company standard.

Individual AI agents are great at one-off tasks, but they lack a shared way to work together on a product over time. Munk AI is the harness runtime for multi-agent collaboration—it brings your agents into a structured local workspace with automated quality gates and human checkpoints, so your team can move work forward while you make the calls that matter.

One Ecosystem, Three Parts

To keep multi-agent collaboration open and portable, the specification, templates, and runtime are cleanly separated:

Site / ProductRoleWhat It Does
open-company.orgOpen Specification (open-company/v0)Defines how to describe a team's goals, products, roles, and schedules in plain files (Companyfile.yaml, Productfile.yaml, AGENTS.md, tasks/*.md, guards/*.mjs).
opencompany.runTeam Template DirectoryHosts battle-tested, open-source agent team setups (such as an open-source software maintenance team) ready for one-click import into Munk AI.
Munk AI (munk.sh)Local Harness RuntimeReads the directory and runs the team on your machine—orchestrating isolated code changes, standing shifts, automated code gates, and human approval checkpoints.

How the Harness Works: Code, Projects, and Gates

Inside Munk AI, your agent team operates across three core pillars:

1. Code: Gated Product Delivery

Turn features and bug fixes into trackable Work Items. Agents clarify requirements, plan, and write code inside isolated Git worktrees. Nothing moves toward delivery until automated Code Gates (builds, typechecks, and tests) pass.

2. Projects: Ongoing Multi-Agent Workflows

Not every job is a source code change. Organize recurring and event-driven work—GitHub triage, package smoke tests, daily releases, research, and content operations—into structured Projects (projects/<id>/tasks/):

  • Zero-Token Guards: High-frequency tasks run a lightweight Node.js script (guards/*.mjs) first, waking an LLM agent only when there is actual work to do.
  • Shared Context & Skills: Agents in a project share files, custom skills (SKILL.md), and shift history.

3. Gates: Human Checkpoints for Critical Calls

Nothing ships or merges behind your back. Local branch merges, pull requests, releases, and sensitive permissions pause at explicit Gates (Waiting) and land in your Inbox until you approve—right from your desktop, via Remote UI on your phone, or by replying directly in Telegram, Feishu, or DingTalk.

Code workbench

Why Local-First?

  • Your Code & Credentials Stay on Your Machine: Munk AI is free for local use. Your source code, Git credentials, and local environment never leave your control. GitHub tokens and write actions are brokered by the local host rather than exposed inside an agent's shell.
  • Plain Files, Zero Lock-In: Team structure, agent boundaries, and schedules live in human-readable YAML and Markdown under the MIT-licensed open-company/v0 spec. Version them with Git or export and share them anytime.
  • Cross-Platform: Native desktop builds for macOS (Apple Silicon & Intel), Windows, and Linux, plus an Android Mobile Host for app previews and mobile approvals.

Next Steps

  1. Getting Started — Install Munk AI, connect a model, and run your first agent team in 5 minutes
  2. Code Track: Gated Product Delivery — See how requirements flow through isolated worktrees, code gates, and safe delivery
  3. Projects Track: Standing Shifts & Multi-Agent Workflows — Run scheduled tasks, zero-token guards, and multi-agent shifts
  4. Open Company Packages: Import & Export — Fork ready-made team templates or package your own setup to share