Zed Rebuilds Claude Agent Integration so Background AI Work Stays Visible
Zed ships Claude Agent natively inside its Rust editor via ACP. Runs at 120 fps with inline comments, live and async review, and context-preserving handoffs. Sign in with an existing Claude Pro or Max subscription, no API key juggling.
Takeaways
- Zed ships Claude Agent natively inside its Rust editor via ACP.
- Runs at 120 fps with inline comments, live and async review, and context-preserving handoffs.
- Sign in with an existing Claude Pro or Max subscription, no API key juggling.
- Built on ACP v2, which adds background work reporting and subagent visibility.
- Codex is next: Zed is updating its ACP adapter to bring it in.
- Zed urges all agent builders to adopt ACP v2 to stay compatible.
Zed has added Claude Agent to its Agent Panel through Agent Client Protocol (ACP) v2. The integration runs inside the editor’s native Rust interface, and Claude Pro or Max subscribers can authenticate with their existing accounts without configuring an API key. Anthropic meters that usage separately from its standard subscription limits.
Where the first adapter lost track
Zed’s earlier beta launched Claude Code as an SDK process and translated its activity through a bridge adapter. Claude Code did not expose ACP directly, so Zed built the translation layer itself. The beta worked for ordinary prompt-and-response sessions but struggled when Claude started subagents or invoked shell commands with run_in_background .
ACP v1 treated a turn as complete when the client received end_turn . It provided no signal for work that Claude Code started independently afterward. As a result, Zed could show a frozen output stream while several subagents or background commands continued running, leaving users without status or progress information.
ACP v2 brings background work into view
Zed rebuilt the integration on ACP v2, which can report independently running work as background activity in each state_update . The editor can keep subagents and shell commands visible after the current prompt-response cycle ends, then expose controls for monitoring or stopping them.
The adapter README describes a broader set of v2 capabilities:
| Area | What ACP v2 exposes |
|---|---|
| Background work | Status updates for subagents and commands that continue outside the foreground turn |
| Session lifecycle | Create, list, resume, replay, fork, delete, and close sessions |
| Execution control | Send prompts, cancel work, request permissions, and ask structured follow-up questions |
| Context management | Additional directories, configuration options, plans, and context compaction |
| Editor surfaces | Terminals, notices, authentication flows, provider metadata, and subagent state |
What changes inside Zed
- Select Claude Agent from the Agent Panel and sign in with an existing Claude Pro or Max account.
- View agent output in Zed’s native interface, which renders at up to 120 fps.
- Attach comments to code lines or messages in an agent thread.
- Review agent work with teammates live or asynchronously through Zed’s collaboration system.
- Hand a task to another agent or person while preserving the conversation and working context.
- Track subagents and background shell commands from the same thread.
Agent threads become shared editor state under this design. Code, review comments, agent output, permissions, and handoffs remain attached to one session, giving teammates a consistent record of the work.
ACP becomes the vendor boundary
Zed is using ACP as the interface between its editor and external agents. Collaboration, permissions, terminals, context compaction, and session controls live in the editor layer, while provider-specific adapters translate those features for each agent. That separation lets Zed improve shared tooling without rebuilding it for every model provider.
Zed is also updating its Codex adapter for ACP v2 and has encouraged other agent developers to migrate. A common v2 interface reduces the editor changes required to add an agent and gives integrations the same session, background-work, and permission primitives.
One login, two meters
Claude Pro or Max sign-in handles authentication without an API key. Anthropic then meters Claude Code activity initiated through ACP against a monthly Agent SDK credit at API rates, separate from the usage limits for Anthropic’s first-party clients. The integration simplifies account setup while retaining that separate billing pool.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
