Skip to content

About

Open-source control plane for coding agents you own — run Codex, Claude Code, Gemini, Cursor and ACP agents side by side with shared history, permissions, handoffs and audit logs.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

4 stars

Watchers

0 watching

Forks

Latest commit

 

History

1,457 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HarnessDesk

Latest release macOS 13 or later MIT licence

Where your agents work, whoever made them.
Run coding agents in Teams. Set the rules with Flows, follow each Run, and step in where the work needs you.

Side by side: Codex and Claude Code each build a self-playing runner game in their own browser, Moonberry Express and Lantern Hop. Codex finishes first, Claude Code follows, the judge picks Codex’s attempt, and Merge waits for you.

Two agents, one brief, a browser each. The judge picks; you merge.

Works with the agents you already run

Codex, Claude Code, Cursor, Gemini CLI, DeepSeek Harness, GitHub Copilot, OpenCode, Cline, Windsurf, Antigravity, Kiro, Kilo Code, Devin, Trae, Amp, CodeBuddy, Goose, Hermes, Junie, Pi, Poolside, Qoder and OpenClaw

…or any agent that speaks ACP. What each one can do →

See what each agent did, step by step

Write, check, review, fix, land: every Run follows the Flow you set. When a check fails or a reviewer asks for a change, you watch it loop back, and you see the one step waiting on you.

The path a Run took: Write, Check passing on its third try, Review asking for one change, Fix, Review approving, Land, and a step waiting for you to merge.

Twelve jobs, one glance

Skip the tour of agent terminals. See what is working, what is in review and what is blocked on you, with the checks, the diff and the pull request on each card.

A Team’s board with columns for Needs you, Working, In review, To do and Ready.

Know what several agents cost you

The Dashboard’s three bands in turn: what each agent spent over 30 days, how much each account has left, and a year of activity. The still shows spending.

Which agent spent what · Before a limit stops you · Where the work went. Docs →

Run many agents without losing track

The Teams page: Teams that need you first, each with its progress, project and Seats.

Every Team, one board
See what needs your answer, what is working and what is ready, across every agent in the Team. Docs →

A Run’s page: the path it took so far, its state and round, and its two Seats.

Follow each step
Watch write, check, review and fix as they happen, and see where a Run waits for you. Docs →

The Flows page: Alignment draws Propose, Align and Build; Side by side draws Competitor, Judge and Referee.

Set the process once
Set the steps, rules and triggers once, such as who checks, who reviews and what lands after both. Docs →

A skill’s page: its SKILL.md, the three agents that load it, its source, its files and its cost per turn.

See what each skill costs
Open a skill from Customize in the sidebar to see who loads it, where it came from and what it costs, then trim the dead weight. Docs →

Runtimes: one needs attention and says why; the ready ones show the account each is signed in as.

Every runtime, ready or not
See which agents are ready, which need attention and why, and which account each signs in with. Docs →

A chat started without a project: the answer, and the folder of its own it was given.

Ask first, pick a folder later
Start a chat before you choose a project: it gets a folder of its own and waits under Chats. Docs →

The same Run, from your terminal

Script it, or let your agent run it. Docs →

An illustrated split screen: a terminal starts the comparison Flow and HarnessDesk follows the same Run to the merge step.

Bring your agents. Keep your setup.

  • No new bill. Each agent keeps its own sign-in, so the plans you already pay for keep working. HarnessDesk is free and adds nothing on top.
  • Your code stays with you. Agents work in your own folders, and the desk keeps its records on your machine. Data boundaries →
  • Read it before you trust it. HarnessDesk is open source under the MIT license. Read the source on GitHub.

Also in the box:

  • Hand-off: pass a conversation to another agent with a packet it can pick up.
  • A browser per Seat: each agent opens, clicks and reads its own pages.
  • Agents: reusable briefs and ceilings, with ten built-in roles to start from.
  • Plugins. Thirteen built in. HarnessDesk projects its 90 built-in plugin tools into each agent’s tool interface. Docs →
  • One policy, one audit log: the same rules, whichever agent asked.

Account, phone and Lead

Settings › Account opens sign-in in your system browser. Approve the named desk and matching code there; return to manage its name, sign out, or delete the account. Agent sign-ins and local work remain separate. The host's requests and stored data are listed in data boundaries.

With local phone access enabled, Settings › Phone shows a code to scan on the same network. Allow the phone on the desk; pairing needs no account link. Settings › Permissions controls Git write steering and its project overrides; the Git record shows gateway writes and observed outside heads.

Lead is an optional conversation for planning work, approving its allowances and following workers, checks and reviews. It needs the startup opt-in HARNESSDESK_LEAD=1 and a runtime whose permission model blocks the network for its child: a standalone host on Node 26, or the desktop app, which carries one. Where no runtime passes, Lead shows as unavailable with the reason; ordinary sessions remain usable. See the interface.

Install

Download the signed, notarized app for macOS 13+, Apple silicon or Intel, from the latest release. It updates itself. Then bring an agent you already use: getting started covers installing and signing in to the main ones, and how to add any other agent from the ACP registry or by its own command.

Build from source with Node 22.19+ and pnpm 10:

pnpm install
pnpm build
pnpm app

CONTRIBUTING.md is the way in; security reports go to SECURITY.md.

Licence

MIT. HarnessDesk is an independent project, not affiliated with, endorsed by, or sponsored by any of the vendors whose agents it drives; their names appear here only to identify their software. The extension kernel is DeepSeek's Cordis (MIT), and the Codex protocol types are generated from openai/codex (Apache-2.0). See THIRD_PARTY_NOTICES.md and TRADEMARKS.md.

About

Open-source control plane for coding agents you own — run Codex, Claude Code, Gemini, Cursor and ACP agents side by side with shared history, permissions, handoffs and audit logs.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

4 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages