Getting Started with Zed ACP Agents

Zed is a fast code editor with a first-class Agent Panel. Beyond Zed’s own agent, you can plug in External Agents through the open Agent Client Protocol (ACP)—so Claude, Codex, Gemini CLI, and others can run as threads inside Zed while keeping their own auth and billing. Editor: zed.dev. External agents docs: External Agents. Agent Panel: Agent Panel. ACP overview for Zed as a client: zed.dev/acp/editor/zed.

For after-work Python, ACP matters because you can try a serious agent without rebuilding your entire toolchain every night. One editor, one practice folder, one reviewed thread.

Who this is for

Professionals who want a lightweight editor and the option to swap agents (Zed Agent vs Claude/Codex/etc.) without learning five UIs. You value diffs, checkpoints, and explicit review after dinner.

See also Top 10 AI agent tools for after-work Python.

Steps to start tonight

1. Install Zed from the official download page

Get the build from zed.dev/download (linked from zed.dev). Prefer the official installer. Sign in if you use Zed account features; keep Privacy / AI settings aligned with practice-only code.

2. Open a tiny practice project

Create evening-python/zed-lab/ with:

# greet.py
def greet(name: str) -> str:
    return f"Hello, {name}"

if __name__ == "__main__":
    print(greet("world"))

Open that folder as the workspace root. Agents follow the project you opened—not your whole disk.

3. Open the Agent Panel and configure a path

Open the Agent Panel (agent: new thread from the Command Palette, or the sparkle control in the status bar). First time: configure either a model for Zed Agent (LLM Providers) or an External Agent via ACP.

For External Agents, open the ACP Registry: Command Palette → zed: acp registry (or Agent Settings → External Agents → Add Agent → Install from Registry). Install one agent you already have credentials for (for example Claude or Codex). Remember: an API key for Zed’s built-in providers does not automatically log you into that External Agent—complete the agent’s own login in its thread.

4. Run one bounded thread and review changes

Start a new thread with your chosen agent. Prompt:

Add greet_many(names: list[str]) -> list[str] to greet.py, keep type hints, print results in __main__. Do not add files or dependencies.

Use Review Changes / keep-reject hunks when the panel surfaces edits. Follow the agent in the editor if you like, but still read the diff. Then:

python greet.py

5. Save a two-minute restart note

In NOTES.md, record which agent you installed, how you authenticated, and what you rejected. Tomorrow should not require rediscovering the registry.

Tiny exercise (15–20 minutes)

Ask the agent for normalize_name(s: str) -> str (strip + title-case) wired into greet, stdlib only. Optional: create AGENTS.md in the project root with one paragraph of conventions (“stdlib only, type hints, no new deps”) so the next ACP thread inherits clearer context.

Pitfalls

• Assuming Zed keys = External Agent keys. ACP agents own auth and billing; configure each once.
• Skipping the registry. Custom agent_servers in settings.json is for advanced/custom agents; start with Install from Registry.
• Huge workspaces. Same after-work rule: tiny lab.
• Blind accept. Use checkpoints / restore when the thread goes sideways.
• Debugging without logs. If an External Agent fails to start, dev: open acp logs helps—then fix auth or PATH before blaming Python.

Why this fits after work

Zed plus ACP lets you sample multiple agents without installing five separate editors. One evening should end with: Zed open on a lab folder, one registry agent authenticated, one reviewed thread, and a NOTES.md of what to reuse. Performance and multi-buffer review help when you are tired; the discipline is still the same—bounded prompts and human acceptance of every hunk.

Next step

ACP agents amplify clear Python taste. SoloQuant’s Intro to Python Programming is sized for after-work lessons—so your threads ask for specific functions, not “clean up everything.”