Skip to content

⚠️ DEPRECATED — AI Collab v0.1 preview

Pre-release
Pre-release

Choose a tag to compare

@shuhongwu shuhongwu released this 17 Aug 15:01
· 43 commits to main since this release

⚠️ DEPRECATED / 已废弃 — do not use. This build cannot start its Harness Host service on end-user machines (unnotarized and/or non-self-contained embedded runtime). Use the latest release instead. / 此版本在用户机上无法启动 Harness Host 服务,请改用最新 release。

AI Collab v0.1 preview

Local multi-agent Scenario harness for macOS. Turn any Git project into
durable task rooms where multiple AI agents work in isolated workspaces and
talk to each other through a typed, audited Host.

Install

  1. Download AICollab.dmg, open it, drag AICollab into Applications.
  2. First launch: right-click the app → Open (the build is signed but not
    notarized yet, so macOS asks once).
  3. iTerm2 is required — participant agents run in iTerm2 windows the Host
    owns and recovers.

First project, from zero

Click Register Project and pick any Git directory. If it has never met
AI Collab, the app offers to draft its declaration files
(project_descriptor.yaml, repo_manifest.yaml, a gate registry, and
starter collaboration templates) from the repositories it finds, validates
the drafts with the same rules registration uses, and registers it in one
step. Existing files are never overwritten.

From there: create a Scenario, and the Host provisions isolated clones of
every declared repository plus a bound Python environment; add participants
(Codex / Claude CLI profiles ship in the box, others via a config overlay)
and they can message each other with the ai-ping transport each
participant receives from the Host.

What's in this preview

  • generic, config-driven project adapter — project identity lives in your
    declaration files, not in code
  • cold-start bootstrap: bare Git directory → registrable project in one click
  • Scenario lifecycle with exact-revision workspace provisioning, drift
    detection, WIP-preserving repair, and fail-closed destroy (no residue,
    but nothing unproven is ever deleted)
  • project unregister from the App sidebar and the CLI
  • participants launch inside the provisioned project checkout, with
    per-participant, Host-issued messaging transport
  • 302 tests, runnable from a bare clone: python3 -m venv .venv && .venv/bin/python -m pip install -e '.[dev]' && .venv/bin/python -m pytest

See the README for the full configuration reference.