23 Jul '25

Running Claude Code within Zed

After several months using Zed’s built-in Agent panel I’ve switched to Claude Code. The main reason is cost. I’m still using the same models—still primarily Sonnet (and occasionally Opus) 4—but using them via the Agent panel incurs per-token billing via the Anthropic API. I was racking up a lot of costs (around 100USD/week) even with a bit of restraint, although still using it fairly heavily during all work days.

Claude Code allows one to pay a flat-fee subscription (I’m now nearly through my first month of MAX 20x, so I’m already ahead). And it’s really nice to have the flat fee; it does change the way you use the agent (and sub-agents).

One downside is that you have to use their claude CLI tool. Since I still want to stay in Zed (even if I can’t use the Agent panel) I’ve added a task which gets me pretty close to the same workflow I had before. Add this to your zed tasks.json:

{
  "label": "claude",
  "command": "claude --dangerously-skip-permissions",
  "reveal": "always",
  "use_new_terminal": true,
  "allow_concurrent_runs": false
}

Anectodally, it also seems like Claude Code is a bit more token-efficient than using the same models via the Zed Agent panel. To my eyes (watching the agent at work) it seems like it’s more parsimonious in only reading sections of files into context, and whole files only when absolutely necessary. But I don’t have hard data to back that up.

And the workflow is really almost as good, with only two things that I’m really missing from my old workflow.

You can’t trivially switch providers/models like you can when you’re using the Zed Agent panel. But the Anthropic models are pretty good—at least equal to best-in-class for the sort of work I do.
Claude code no longer has direct access to the LSP diagnostics—because that stuff in general Just Works^TM in Zed, it was nice to not have to set up extra MCPs to get access to the language-aware tooling stuff. I used to be able to just say “fix the warnings in this project” and it’d do what I meant, but now I have to either set up a MCP server, or hope that Claude (possibly with a hint in CLAUDE.md) knows how to run the CLI tools to get the same information.

The thing that puts me at ease with this sort of (big) workflow chnage is that there’s really not too much lock-in; if the pricing landscape changes in the future I’d happily switch to another provider. The main skill—context management, and figuring out how to communicate to an LLM what needs to be done in clear, concise language—is unchanged.

If this helps you get the most out of your setup, then happy Claude-in-Zedding.