The Collaborative Agent Platform
Dispatch, coordinate, and manage AI bots across your projects.

---

Collabot is home base for your AI bots. It runs as a persistent service, dispatching bots to work on tasks across multiple projects. You define the projects, roles, and bots. Collabot handles the rest — dispatch, coordination, event capture, context reconstruction, and communication.

A project is a logical product that may span multiple repositories. Collabot provides the infrastructure; projects bring the domain knowledge.

How it works

The harness is the core — a persistent Node.js/TypeScript process that manages everything. Interfaces connect to it through adapters. No adapter is primary. The harness runs with or without any of them.

                  +------------+
       Slack ---->|            |
        CLI ----->|  Harness   |-----------> Bots
  WebSocket ----->|            |               |
        TUI ----->|            |         +-----------+
                  +------------+         |  Project  |
                                         |   repos   |
                                         +-----------+

Adapters connect humans to the harness — DM a Slack bot, type in a TUI, send a WebSocket message, or fire a CLI one-liner. The harness resolves the right project, role, and bot, then dispatches.

Roles define behavioral profiles — identity, model preference, and permissions. They're tech-stack-focused, not project-specific. A .NET developer role works on any .NET project.

Bots are persistent identities that sit above roles. A bot has personality, a default project, a role assignment, and (eventually) memories. Think of roles as the job description and bots as the person filling it.

Current state

Collabot is pre-release (0.x) and under active development. What's working today:

• Multi-project orchestration — register projects, dispatch bots with role-based context
• 4 adapters — Slack (multi-bot, per-bot Socket Mode), CLI, WebSocket (JSON-RPC 2.0), TUI (.NET 10)
• Bot session pattern — persistent bot identities with resume-per-message SDK sessions
• Event system v2 — canonical event streams scoped to dispatches, full agent lifecycle capture
• Multi-bot coordination — PM bots dispatch teammates via MCP tools, await results, synthesize
• Task lifecycle — create, dispatch, capture, close — all scoped to projects
• Parallel dispatch — git worktree isolation for concurrent bots on the same repo
• 376+ tests, TypeScript strict mode, CI on every PR

Quick start

npm install -g @anthropic-ai/claude-code   # prerequisite — the AI engine
claude                                      # log in (browser OAuth, API key, or cloud provider)

npm install -g collabot
collabot init          # scaffold ~/.collabot/
collabot setup         # verify CLI, pick roles and bots
collabot start         # start the harness

See the package README for configuration details.

Vision

We're building a team, not a workflow tool. Collabot is where your bots live — they have roles, personalities, projects they care about, and (eventually) memories of past work. The platform ships infrastructure, not content. A fresh install is essentially empty — you build up the team that fits your projects.

Three principles drive the design:

• Curated context over large context. The right 10K tokens beats 100K tokens of vibes. Collabot assembles precise context, not maximum context.
• Documentation is memory. Bots have no memory between sessions. What's written down is what they know. Docs evolve in the same commit as code.
• Iterative formalization. Start loose, use the system, let the shape emerge, then codify. The loose version isn't debt — it's research.

Documentation

Document	Description
Architecture	Platform architecture
Vision	Origin story, design philosophy, growth model

License

MIT