Zed Agent
Use Zed's built-in agent panel as the project agent — no external CLI, no extra harness, just the editor.
The Zed Agent is the default agent that ships with the Zed editor. When you pick it as a project's runtime, Helix doesn't launch a separate CLI in the sandbox — the user opens Zed's existing agent panel and interacts with it directly. This is the lowest-friction option: no Anthropic CLI to install, no ACP bridge to maintain, no extra config to keep in sync.
Pick the Zed Agent when you want a fully in-editor experience and you're happy to use whichever model providers you've configured in Zed itself.
Selecting Zed Agent for a project
Set the assistant's runtime to zed_agent in your project YAML:
apiVersion: helix.ml/v1alpha1
kind: Project
metadata:
name: my-app
spec:
description: "App using Zed's built-in agent"
repositories:
- url: "https://github.com/org/my-app"
branch: main
primary: true
agent:
name: "Zed Agent"
runtime: zed_agent
model: claude-sonnet-4-6
provider: anthropicLike the other runtimes, provider resolves through Helix's /v1 proxy in OpenAI-compatible mode — so any provider you've configured under Account → AI Providers is available, whether that's Anthropic, OpenAI, Together, or a local runner.
How the model is wired up
Helix injects Zed's language_models and agent settings into the sandbox so the editor talks to Helix's /v1 proxy for the selected model. The agent panel works as it does outside of Helix — the streaming-thread UX, in-editor diff approvals, and built-in tool calls all behave the same way.
What you give up versus a dedicated agent CLI:
- No vendor-tuned harness. You get Zed's generic agent plumbing, not Anthropic's hand-tuned Claude Code prompts or Block's recipe model.
- No project-scoped recipes. Recipes (see Goose) are a Goose feature, not a Zed one.
- No subscription auth. The Zed Agent always goes through API-key routing via the configured provider.
When to pick Zed Agent
- You want the simplest possible setup — open the editor, talk to the agent, no extras.
- You're using Zed outside of Helix already and want the same experience inside a sandbox.
- You don't need recipes, baked task parameters, or vendor-specific harnesses.
See Claude Code for Anthropic's official CLI agent, Goose for project-scoped recipes, or Qwen Code for an OpenAI-compatible open-source harness.