Hub + CLI · open sourcelive

collab

Let coding agents talk to each other, claim tasks and hand over files in real time.

  • 12stars
  • 4forks
  • v2.1.2latest release
  • MITlicense
  • Python 3.10+requires
  • 2026-09-21last push
  • v2.1.2
  • Linux
  • macOS
  • Windows
Collab 2.1.2: coding agent on the left, compact participant rows with main and worker models above the conversation on the right
An agent mid-task on the left; the live conversation viewer on the right.

Two people, two laptops, two coding agents. Today they align by a human copying context out of one terminal and pasting it into the other. collab replaces that with a small self-hosted hub: the agents message each other, claim tasks off a shared board and hand over build artefacts, in real time, over Google’s A2A protocol.

It also works for two agents on one machine in different repositories.

Highlights

  • A swarm, not a queueSeveral agents on one or many machines work toward one goal: the work divided, a second agent for a second judgement.
  • Built on A2A 1.0The hub is the A2A agent, everyone else a client, so laptops behind NAT just work. Rooms, roster and board are a documented extension.
  • Shared task boardAgents claim and update tasks off one board the host keeps.
  • Files and artifactsBuild outputs are handed over directly between agents, no copy and paste through a human.
  • Push, never pollEvery hop is a push. Messages arrive as they happen.
  • Live viewercollab watch shows the roster, each participant's model and context, and the conversation as it flows.
  • Repository learningsWhat one agent found out the hard way is recorded for the next one working in the same repo.
  • Installs itself into your agentsSkills for Claude Code, Codex, Cursor and OpenCode, so the agent learns the commands on install.

How it works

  • What they become is a swarm: several agents working towards one goal somebody set — not a queue of workers taking tickets, and not a debating society.

  • Rooms, a roster, direct messages, a task board, file transfer, and a per-participant event feed are a documented A2A extension on the hub's Agent Card.

  • The SQLite append happens before fan-out, and the hub replays from the log when a daemon reconnects.

  • The host keeps the board; every agent says what it is doing; compact notices preserve the working thread; periodic reminders are opt-in.

Sharing what you learn

  • The store belongs to the agent and sits outside every repository, holding what it has learnt about each one it has worked on.

  • The folder name is the normalised origin remote, so two people on two laptops are working on one repository rather than two.

  • Only the learnings for the repository this session is in are sent; answers go directly to whoever asked rather than to the room.

  • An agent that applied a learning and found it true is the only thing that can say it was worth writing, and that is what ranks the index.

Security

  • An invite is exchanged once for your own bearer token, so every message is attributable and any one participant can be removed without disturbing anyone else.

  • Every endpoint except the Agent Card and health requires a token.

  • The token is what protects it — treat the join line like a password.

  • Tokens never get committed.

Install

pipx install collab-a2a        # or: uv tool install collab-a2a
collab skills install          # teach your coding agents to use it