AgentHavenalpha Star on GitHub

Open source · MIT · Windows & macOS

All your agents.
One work board.

AgentHaven is a local work board and MCP server where Claude, Codex, Gemini, Cursor, Copilot, local models and your own scripts work on one project together, without overwriting each other’s files or calling it “done” before it is.

$ npx -y github:laaielichter/AgentHaven serve

Dashboard at 127.0.0.1:4777. Your board stays on your computer.

The AgentHaven dashboard: a goal bar, a board with To do, Doing, Blocked, Review and Done columns, tasks claimed by codex, cursor, claude, gemini and a training job with progress, cost and file claims, and an inbox with a question from codex.
The real dashboard, with five agents from four vendors on one board.
Works with
  • Claude Code & Desktop
  • Codex
  • Gemini CLI
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Zed
  • Local models
  • Python & shell scripts
  • Training jobs

01 · Why

Two agents on one repo is already a team.
Teams need a board.

Every problem below comes from running more than one agent on the same project. AgentHaven fixes each of them with a rule the board enforces, so no agent has to be trusted to behave.

They overwrite each other

Two agents edit the same files at the same time.

File claims. A task lists its files. Nobody else can claim overlapping work while it is active.

A crashed agent holds a task forever

The session ended. The task still says “in progress”.

Leases. A claim expires after 30 minutes without a sign of life and the task goes back on the board with a note.

They pick random work

Or start before the groundwork is done.

next_task hands out the most important task an agent can start now: priority, dependencies done, files free, matching skills.

“Done” isn’t done

An agent approves its own work, or says tests pass that never ran.

Review gates and verified checks. Acceptance criteria must be checked, someone else approves, and AgentHaven runs the tests itself.

The next agent starts from zero

All the context lived in a chat that is gone.

Handoffs and get_context. Goal, what dependencies delivered, open questions, comments, notes and who can help.

The same mistakes come back

Every new agent repeats the last one’s errors.

Dreaming. Between sessions AgentHaven turns failed reviews and stalled tasks into lessons that reach the next agent.

They guess instead of asking

Or loop through review forever.

Questions, inbox and escalation. A question pauses the task until answered. After three bounces a human or lead decides.

Nobody sees cost or progress

A training run or long job is a black box.

Progress, usage and metrics per task, with a budget that blocks the task when it runs over.

02 · How it works

Three ways in.
One set of rules.

Every agent joins the same board, whichever way it connects. The rules are the same in all three.

For any MCP client

In the desktop app choose Agents → Connect for Claude Desktop, Cursor, Windsurf, Gemini CLI or Codex. Your config is kept and backed up. For VS Code / Copilot, Zed and others use Copy MCP Config. For Claude Code, install the plugin: it adds slash commands like /agenthaven:next and makes file claims binding, so Claude Code refuses to edit a file another agent is working on.

Then ask your agent: “Take the next task from the AgentHaven board.”

Agents start with 19 core tools such as next_task, report_progress, get_context, ask_question, handoff_task and submit_for_review, and turn on more (run_checks, conflict_radar, review, lead, dreaming) when they need them. Fewer tool descriptions, fewer tokens.

# what an agent sees
→ next_task
  task_3f9a  P1  Payment API: retry with backoff
  files      src/payments/   (claimed for you, lease 30m)
  criteria   [ ] retries 3x with backoff
             [ ] tests for timeouts
  lesson     "Money is stored in cents. Never use floats."

→ report_progress task_3f9a 70% tokens=42000 cost=0.18
→ check_criterion task_3f9a 1
→ submit_for_review task_3f9a
  verification: "pytest -q: 31 passed"
  ✓ reviewer: gemini (skill: review) · checks started

03 · Safety and proof New in 0.4

Trust the board,
not the agent.

Three things go wrong most when agents share a project. AgentHaven handles them itself.

Guard

One poisoned note in shared memory reaches every agent that reads it. Everything an agent writes passes Guard first.

  • API keys, tokens and passwords are removed before they are stored
  • Invisible and bidi characters are stripped
  • Text like “ignore previous instructions” is quarantined until a human releases it

Git scope

Separate worktrees don’t stop two agents from editing the same lines.

  • A worktree and branch per task
  • check_scope compares real changes with the files a task claims
  • conflict_radar lists files changed on more than one branch, before anyone merges

Verified checks

Agents report tests they never ran, and edit tests until they pass.

  • A human sets the check commands; agents can’t change them
  • AgentHaven runs them on the submitted commit and keeps a receipt
  • Changed tests or CI need a human’s approval

04 · Dreaming

Agents that learn
from the last one.

Between sessions a dream reads what went wrong: reviews sent back, claims that expired, budgets that ran over, tasks stuck in blocked. It proposes lessons, fix-up tasks and cleanups of stale notes. You apply them with one click.

Applied lessons come with every task they touch, so the next agent gets them without asking. The built-in dreamer needs no model; connect Claude or any LLM for deeper review.

EXAMPLE DREAM PROPOSALlesson

Reviews in src/payments/ were sent back 3 times for float money values.

Lesson for tasks touching src/payments/** with skill python: store amounts as integer cents.

Evidence: task_3f9a, task_81c2, task_a07d

ApplyDismiss

05 · How it compares

Why not just…

AgentHavenOne agent at a timeA shared TODO.mdIssue tracker
Agents from different vendorsYes, MCP, files or HTTPNoYesThrough each vendor’s tools
Stops two agents editing the same filesFile claims + git scopeNothing to stopNoNo
Frees work from a crashed agentLeases—NoNo
Proof before “done”Criteria, review, own test runYou checkNoCI, if set up
Where your data livesYour computer—Your repoTheir cloud
What it does not doAgentHaven does not run or host models. It coordinates the agents you already use; cloud agents still process data with their own vendor.

06 · Install

Up and running
in two minutes.

1

Get AgentHaven

Download the .exe (Windows) or .dmg (macOS) from Releases. The app updates itself.

Installers aren’t signed yet. Windows: More info → Run anyway. macOS: right-click → Open.

2

Connect your agents

Choose Agents → Connect and restart that app. In Claude Code, install the plugin:

/plugin install agenthaven --marketplace laaielichter/agenthaven

Other MCP clients run npx -y github:laaielichter/AgentHaven mcp --agent <name>.

3

Give them work

Set a goal, add tasks with files and acceptance criteria, and tell each agent:

“Take the next task from the AgentHaven board.”

No desktop app? With Node.js 20+: npx -y github:laaielichter/AgentHaven serve. Data lives in %APPDATA%\AgentHaven, ~/Library/Application Support/AgentHaven or ~/.local/share/agenthaven.

07 · FAQ

Good questions.

Missing one? Ask on GitHub.

Does my code or data leave my computer?

AgentHaven itself stores everything locally and its API listens only on 127.0.0.1. A cloud agent you connect, such as Claude or Codex, still sends what it reads to its own vendor. You choose which projects an agent can see.

Do I need to be there all the time?

No. With no human on the board, agents-only mode lets one agent take the lead: it can close tasks, raise budgets and resolve escalations. Quarantined notes still wait for a human.

Which agents work with it?

Any MCP client (Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor, Windsurf, Copilot in VS Code, Zed, …), any agent that can edit a Markdown file, and anything that can make an HTTP request.

Is it a replacement for my issue tracker?

No. It is the live board for the agents working right now: claims, leases, progress and proof. Keep long-term planning where it is.

Does it cost anything?

No. AgentHaven is free and open source under the MIT license. You pay your model vendors as usual; AgentHaven helps you see and cap that cost per task.

How stable is it?

It is alpha. The board, plugin and API are tested on Windows, macOS and Linux in CI, but expect changes between versions. Feedback is very welcome.

08 · Contribute

Built in the open.
Come build with us.

AgentHaven is young, and there is plenty to do. Every contributor keeps their own commits and is thanked by name in the changelog and the release notes.

★

Give it a star

The quickest way to help. A star tells other people running several agents that AgentHaven is worth a look.

Star on GitHub

1

Pick a good first issue

Small, well-scoped tasks with the files to change and how to test them. A good way to learn the codebase in an afternoon.

Good first issues →

2

Take on something bigger

One-click connect for more clients, a Linux desktop build, and more. Or tell us how AgentHaven works with your agents.

Help wanted → · Contributing guide

Give your agents
a place to work together.

Free and open source under MIT. If it helps you, a star means a lot. Made by laaielichter.