Claude Agent

Beta: Claude Agent is in beta. Its behavior can change between releases.

Use Claude Agent in a Delta thread with your existing Claude subscription or Anthropic Console account. Delta connects to it through the Agent Client Protocol (ACP).

Claude Agent is built on Anthropic’s Claude Agent SDK, so it loads your skills, commands, and memory the same way Claude Code does. Delta also passes the MCP servers you’ve connected through to it.

Claude Agent is a separate agent, not a model for Delta’s own agent. A thread works the same way with either one. You can inspect Claude’s changes, comment anywhere in the code or conversation, and review the work with teammates. You can also hand a thread from one agent to the other and keep its context.

Before you start

The Delta desktop app. Claude Agent turns run on your computer. Delta on the web and Cloud Runners can show Claude Agent turns but can’t run them. See Limitations.
A Claude subscription or an Anthropic Console account. See Sign in.

Set up Claude Agent

Turn it on

Choose Claude Agent when Delta’s onboarding asks how you want to use models. Or, at any time:

Open Settings > LLM Providers.
Under Agent Connections, turn on Claude Agent Integration.

Delta installs Claude Agent when you choose it during onboarding, when you open Settings > LLM Providers with it turned on, or when you first send a message with it. Opening a thread or choosing Claude Agent in the model picker doesn’t install anything. The first installation can take a few minutes.

Claude Agent uses its own login, separate from the providers in Settings > LLM Providers. Choose one of these:

Claude subscription. Usage counts against your Claude plan’s limits.
Anthropic Console account. Usage is billed to your Anthropic API account.

In both cases, Claude Agent usage isn’t metered against your Zed plan.

If you’re signed in to Claude Code on this computer, Claude Agent uses that login. Otherwise, sign in during onboarding, with Sign In to Claude Agent under Agent Connections in Settings > LLM Providers, or from the sign-in options the message input shows when you send. Delta runs the login on your computer, and nothing you enter is saved to the thread or shared with collaborators.

Agent Connections shows the signed-in account. If you sign in or out outside Delta, choose Refresh there to update it.

Note: An Anthropic API key in Settings > LLM Providers applies only to Delta’s own agent, not Claude Agent. Delta also ignores an ANTHROPIC_API_KEY exported in your environment, so a key you set for other tools doesn’t switch a subscription login to API billing.

Start a thread with Claude Agent

In a thread, open the model picker beside the message input.
Under Other Agents, choose Claude Agent.
Choose a model from the list beside it.
Send your message.

If Delta hasn’t connected to Claude Agent on this computer yet, the model list is empty. Send a message with Claude Agent selected, and Claude reports its models once it connects.

Controls beside the message input show the options Claude offers, such as thinking effort and mode. The available modes come from Claude. Delta remembers your choices as defaults for new Claude Agent threads.

What appears in the thread

Claude Agent turns show the same kinds of content as Delta’s agent:

Messages and thinking summaries
Tool calls, file diffs, and terminal output
Plans and to-do lists
Usage and context compaction
Notices from Claude

Turns show Claude’s name and icon. You can review and sync Claude’s changes the same way as Delta’s agent changes. See Review & Sync Changes.

Type / in the message input to use Claude’s slash commands.

You can fork a Claude Agent thread, edit and resend an earlier message, start a review, or Land changes, the same as with Delta’s agent. After you edit an earlier message, Claude starts a new session and receives the conversation as context.

Claude’s subagents

When Claude starts its own subagents, each one appears as a read-only child thread labeled Run by Claude Agent, with its own status and a Stop button. These are Claude’s subagents, not Delta’s subagent profiles.

Approve actions and answer questions

Claude Agent asks before taking actions that need permission, such as running a command or editing files, depending on its mode. It can also ask you a question or ask you to open a link.

Delta shows every option Claude offers, including an option to always allow an action. Delta never approves a request automatically. A request waits until someone answers it.

Requests go to the person whose computer runs the turn. Other collaborators can see the request but can’t answer it.

Switch between Claude Agent and Delta’s agent

You can switch agents at any point in a thread with the model picker. Stop a running turn before you switch.

Delta passes the turns an agent didn’t see to the new agent as context. When you switch back to Claude Agent, Claude continues its own session with its full context. Turns from Delta’s agent reach Claude as Delta’s record of the conversation, not Delta’s agent’s internal state.

Switching a thread’s agent changes it for everyone in the thread.

Collaborate on a Claude Agent thread

Everyone in a shared thread sees Claude Agent’s output. Each computer keeps its own Claude session:

A turn runs on the computer of the person who sends it, with their Claude login.
Each collaborator who sends Claude Agent turns needs Claude Agent turned on and signed in on their own computer.
When a collaborator sends after turns that ran elsewhere, Delta passes those turns to their session as context.

See Collaborate in a Thread.

Tools available to Claude Agent

Claude Agent has its own built-in tools for reading, editing, searching, and running commands. Delta also offers it:

Your connected MCP servers. Servers in Settings > MCP that show as connected. Delta relays them through your Delta account, so Claude doesn’t need to sign in to them, and Delta’s credentials never reach Claude. Servers that still need authorization aren’t offered. See MCP Servers.
The Delta CLI. Claude can use the Delta CLI to read and message your other threads.

Retry a failed or interrupted turn

If a turn fails or is interrupted, for example because you quit Delta mid-turn, Delta keeps any partial output and offers Retry. Delta never replays a turn on its own. If Claude may already have acted on the message, the thread warns that Retry can repeat those actions.

Thread titles and summaries

Claude names an untitled thread. Your renames always take priority, and Regenerate Thread Title uses your Delta model.

Delta’s own model doesn’t write search, turn, or review summaries for Claude Agent threads by default, because that would send their content to a second provider. To allow it, turn on Summarize Claude Agent and Other Agent Threads under ACP Agents in Settings > Thread. This may add usage charges from that provider.

Sign out

In Settings > LLM Providers, under Agent Connections, choose Sign Out, then confirm.

Important: Signing out of Claude Agent also signs out Claude Code and other apps that use the same login on this computer.

Turning off Claude Agent Integration prevents new Claude Agent turns but keeps you signed in. It doesn’t stop a turn that’s already running.

Limitations

Desktop only. Delta on the web and Cloud Runners show Claude Agent turns but can’t run them. In those places, the thread offers Switch to Delta to continue with Delta’s agent.
Claude is the only supported agent. Other ACP agents and custom agent commands aren’t supported at this time.
Not available in subagent profiles. Delta’s Worker, Scout, and Reviewer profiles use Delta’s agent.

Models & Providers covers models for Delta’s own agent.
MCP Servers covers connecting tools both agents can use.
Collaborate in a Thread covers sharing threads with your team.
AI Privacy & Telemetry covers what each agent sends to its provider.