Overview
Ace is an open-source AI coding workspace built around a simple observation: most AI coding tools lock you into one model, one session, and one working directory. Real engineering work doesn't look like that. You want to run several agents in parallel, compare providers, keep each experiment isolated, and watch everything happen live.
Ace treats the workspace — not the chat window — as the unit of work.
Architecture
The system is split into three planes:
- Orchestration plane. A provider-agnostic session manager that normalizes OpenAI, Anthropic, and local model APIs behind a single streaming interface. Sessions are resumable state machines, not ephemeral request loops — every tool call, diff, and message is an event in an append-only session log.
- Execution plane. Each agent session gets an isolated git worktree. Agents can edit, build, and test aggressively without touching the primary branch, and results merge back through normal git review. Worktree lifecycle (create, prune, recover) is fully managed.
- Surface plane. A real-time UI fed over WebSockets. Because the session log is the source of truth, any client can attach mid-session and replay to the current state — the same mechanism powers live streaming, reconnects, and session sharing.
Key decisions
Event-sourced sessions. Persisting the event log instead of chat transcripts makes resume, replay, branching ("fork this session"), and audit trivial. It also decouples the UI from any provider's message format.
Worktrees over containers. Containers give stronger isolation but slower feedback loops. Worktrees keep the agent inside the developer's actual toolchain — same compilers, same caches — while still isolating changes at the branch level.
Browser as a first-class tool. Agents can drive a real browser for docs lookup, UI verification, and end-to-end checks, with screenshots streamed back into the session timeline.
Status
Actively developed. The core workspace, multi-provider orchestration, worktree isolation, and streaming UI are functional; current work focuses on multi-agent coordination policies and richer replay tooling.