AWP vs ACP

The Agent Client Protocol (ACP) standardizes communication between code editors and coding agents, the way LSP standardized language servers. AWP borrows deliberately from ACP’s playbook — schema-first artifacts, capability-gated features, per-feature RFD stabilization — while covering a different boundary.

	ACP	AWP
Boundary	editor ↔ coding agent	agent ↔ world
The “other side” is	a mind (the agent) serving a user in an editor	an environment enforcing physics and safety
Core exchange	prompt turns, tool calls, diffs, plans	observations in, typed intents out
Time	conversational turns	lockstep ticks or wall-clock streaming
Data	text/JSON content blocks (Markdown default)	multi-modal binary streams + JSON control
Human-in-the-loop	permission requests, elicitation	approval flow, takeover, shadow sessions
Default transport	stdio subprocess	local WebSocket, stdio optional

Both protocols assume the two parties share an environment: ACP spawns the agent beside the editor; an AWP agent connects to a world running beside it. The tell for which protocol you need: in ACP both parties are software minds cooperating on a codebase; in AWP one party is a world that the other inhabits. They compose. An ACP coding agent can hold an AWP session to test the software it writes inside a simulated environment, and an AWP GUI world can host an editor an ACP agent drives. Where AWP carries human-readable content (approval descriptions, event summaries), it reuses MCP/ACP-style content blocks with Markdown as the default text format, so clients can render both protocols with one component set.

Get started

Core concepts

Guides

Compare