Name: acowork
Author: acowork

One platform for humans & AI agents to ship work together

Collaborate with multiple agents — each with private memory, real-time context awareness, and tool execution.

Full-stack development framework with iterative debugging, token tracking, and context snapshots for deep insight into AI reasoning.

✨ Highlights

ACowork is a multi-user, distributed, AI-native collaboration platform. People, AI agents, projects, and
documents share one runtime — so a team can plan work, write specs, review code, and ship features alongside
their agents the same way they would alongside a human teammate.


👥 Multi-user, not single-actor	Real accounts, roles, sessions, per-user audit. Humans and agents share the same chat / project / doc surface — no "AI side panel" silo.
🌐 Distributed agents, one control plane	Gateway brokers MQTT + HTTP reverse proxy; Node Agents run on every host that hosts Runtimes (GPU box / workstation / cloud VM) — single-host and multi-host share one protocol path.
🗂️ AI-native project & document management	`acowork-pm` and `acowork-doc` are first-class services; agents read, write, and collaborate on tasks and documents through REST + MCP — including real-time multi-user editing (Yjs CRDT).
🚀 Standalone or cluster — same binary	Start a single host for personal use; flip a flag and the same binary becomes a cluster across your team. No local-vs-remote code branches in Gateway.
🧩 Two-stage sessions	Chat to shape the plan, Execute to run it end-to-end. The chat stage holds no execution tools — it can only talk, read, and hand off.
⚙️ Universal protocol	One contract for runtimes, nodes, and devices. PC, phone, pad, or embedded devices — anything that speaks it joins the network.
🔒 Process-level isolation	Every Agent runs as an independent OS process with its own filesystem, memory database, and sandboxed tool execution.
🧠 Biomimetic memory	Layered memory on a per-Agent SQLite store — episodic / knowledge / procedural / autobiographical, retrieved by vector + FTS5 hybrid search.
🛡️ Multi-tier permission system	Admin/user roles · per-node, per-agent, and per-session visibility with explicit user grants. Least privilege, by default.
🛠️ Full-stack dev loop	Desktop App (Tauri v2) supports DevMode: conversational debug, skill hot-reload, step & resume, publishing wizard.

Every Agent is still an independent "digital being": its own runtime process, private memory, workspace, and
configuration. Tune prompt, tools, and memory = build an agent. Personal and sensitive data is
automatically stripped during packaging, so you can share an agent's capabilities freely without leaking your
private memories.

🏛️ Architecture

Humans drive it from Desktop or Mobile; agents and domain services reach it over the same protocol path. Under the
Gateway sits a flat three-tier chain: Gateway → Node Agent → Runtime.

Tier	Count	Responsibility
Clients · Desktop App · Mobile App	per device	Tauri v2 UI over HTTP + MQTT — same surface, same account
Gateway · `acowork-gateway`	1 per cluster	HTTP `:19876`, MQTT `:19875`, reverse proxy, package manager, intent routing, global resources
Node Agent · `acowork-node`	1 per host	process table, local package store, reverse proxy `:19900`, LSP supervisor
Runtime · `acowork-runtime`	1 per `.agent`	loads the package, LLM loop, tool execution, loopback HTTP

Four co-resident services sit on the same path, exposed to humans and agents alike over REST + MCP:
acowork-user (accounts · roles) · acowork-pm (projects · tasks) · acowork-doc (document library) ·
acowork-memory (memory service).

Standalone — the Gateway auto-spawns one local Node. Cluster — enroll acowork-node on every host.
One binary, one protocol path, zero local-vs-remote branches in the Gateway.

ACowork System Architecture

Full topology, protocol paths, and the frozen design boundaries:
docs/design/zh/01-overview.md ·
docs/module-design/zh/ ·
docs/adr/zh/

🚀 Quick Start

Install the Desktop App for your platform — Windows .msi, macOS .dmg, Linux .deb / .rpm /
AppImage. The installer bundles the Gateway and every service it supervises.

On first launch, pick one of three Gateway modes.

1. Local

The app starts a Gateway for you.

Step 1 — launch the Desktop App.
Step 2 — choose Local in the Gateway mode picker.

2. Remote

Connect to a Gateway running on another machine.

Step 1 — on the Gateway machine, start it:

acowork-gateway --daemon

Step 2 — launch the Desktop App.
Step 3 — choose Remote and enter http://<gateway-host>:19876.

3. Relay

Reach a Gateway through the cloud relay.

Step 1 — on the Gateway machine, enable the tunnel in gateway.toml
(~/.acowork/acowork-gateway/config/):

[relay]
enabled = true
url = "wss://<relay-domain>/tunnel"

Step 2 — launch the Desktop App.
Step 3 — choose Relay and enter https://<gw-id>.<relay-domain>.
The <gw-id> is shown in the Desktop App's Gateway settings.

Agents are created through the Create Agent wizard in the app.

🔨 Building from Source

Development build — native window with a hot-reloading frontend:

cd apps/acowork-desktop
npm install
npm run core:build:debug      # build the core binaries the app shells out to
npm run tauri dev             # native window, HMR frontend

Release installers (.msi / .dmg / .deb / .rpm / AppImage):

cd apps/acowork-desktop
npm run tauri build           # beforeBuildCommand auto-runs core:build:release + npm build + tauri bundle

Full build, packaging, CI, and deployment guide: dev/README.md · AGENTS.md

📚 Documentation

Architecture design: docs/design/zh/
AI-generated code wiki (DeepWiki): deepwiki.com/acowork/acowork
Module-level design: docs/module-design/zh/
Design notes & decision history: docs/adr/zh/
Developer conventions: AGENTS.md

🧪 References & Acknowledgments

ACowork's design draws inspiration from:

ZeroClaw 🦀 — trait-driven runtime, security decorators, streaming parser
Grafeo — hybrid (vector + lexical) retrieval over a local store; the storage engine has since moved to SQLite (ADR-082)
Mem0 — multi-level memory, user/session/Agent state
HippoRAG — neurobiology-inspired long-term memory, associative diffusion
LightMem — lightweight memory compression
OpenCode — multi-agent collaboration, provider-agnostic design

🤝 Contributing

The project is in active implementation (Alpha). Code, design feedback, and reviews are all welcome:

Open issues for bug reports, proposals, or design feedback
Read AGENTS.md for project conventions before opening a PR
Read CONTRIBUTING.md for the build, test, and review gates
Report security issues privately, per SECURITY.md
Participation is governed by the CODE_OF_CONDUCT.md

📄 License

Apache-2.0 — see LICENSE for details.

ACowork — AI Agents Collaboration Platform
An all-in-one platform for humans and AI agents.