The open-source collaborative IDE where you and your friends — human and AI — code together in one shared, live workspace.
ICE is a self-hostable collaborative coding environment built around four pillars:
- 🤖 Chat with AI agents — talk to a coding agent (or several) that can read, write, and run code inside the shared workspace, not in an isolated sandbox.
- 🎙️ Talk to your team — voice rooms, group chat, and DMs, built in. No second app.
- 👥 Code together, live — real-time collaborative editing (CRDT-backed) with presence, shared cursors, and shared terminals.
- 🪐 Own your stack — one self-hostable server that ships with its own git server. Your code, your infra, forever.
The thesis: most "multi-agent coding" tools isolate agents in separate git worktrees and make humans merge afterward. ICE does the opposite — humans and agents share one live workspace, the way Figma made design multiplayer. See docs/VISION.md.
All six releases are shipped, tagged, and verified end-to-end:
| Release | Theme | What it added |
|---|---|---|
| v0.1 | Two friends, one live file | CodeMirror 6 + Yjs live multiplayer editing, named cursors, durable rooms (SQLite), auth + invite links, one-command self-host |
| v0.3 | Git comes home | Bundled smart-HTTP git server (git clone your room) + the CRDT↔git convergence seam (commits projected from the doc), file tree + source control |
| v0.5 | Your first AI teammate | An AI agent via the ACP↔Yjs adapter — edits through the doc (attributable, undoable), an approval gate, and a shared terminal |
| v1.0 | Code with friends, out loud | LiveKit voice + group/DM chat + in-UI git + one-click agent rollback + preview + hardening (OIDC/Postgres/roles, self-host docs) |
| v1.5 | A squad of agents | Orchestrator routing subtasks to specialist roles, multiple agents co-editing concurrently, per-agent policies, MCP palette, opt-in snapshots |
| v2.0 | Scale out and integrate | Redis horizontal scale-out, audit log, optional Forgejo + Helm chart, governance + RFC + threat model, community extension surface |
Verification: ~198 server tests + per-release e2e checks (apps/web/scripts/*-check.mjs), pnpm typecheck/build clean, docker compose config valid.
Decisions: docs/DECISIONS.md · Roadmap: docs/ROADMAP.md · Per-release specs: plans/releases/ · Resuming work? docs/AGENT_HANDOFF.md · Live log: PROGRESS.md.
Some capabilities need a real external service to exercise at runtime (a LiveKit SFU for audio, an IdP for OIDC, Postgres, a Redis cluster, a real ACP agent binary + key). These are wired + unit-tested and bring-your-own at deploy time; CI verification uses a built-in mock agent and the single-node defaults.
apps/
web/ # Browser client (editor, chat, voice, agent panel)
server/ # Realtime + API + AI agent host + git server
packages/
shared/ # Shared types & protocol definitions
docs/ # Vision, competition, architecture, roadmap, decisions, handoff
plans/releases/ # One spec per release (v0.1 … v2.0)
PROGRESS.md # Running build log (updated every tick)
Target experience (not yet wired up):
docker compose up→ openhttp://localhost:3000→ create a room → invite friends and agents.
Apache-2.0 — one permissive license across the whole repo, with DCO (no CLA). Forkable forever, self-host forever. See LICENSE, CONTRIBUTING.md, and GOVERNANCE.md.