Run Zed on Unlimited Claude

Zed's assistant is quick to reach for — inline edits, refactors, quick questions — which means it accumulates calls all day without you noticing. A flat-rate unlimited Claude plan suits that rhythm perfectly: you use the assistant freely and the cost never changes.

The provider that takes a custom URL

To put an unlimited gateway behind Zed you need a provider that accepts a base URL, and the native Anthropic provider does not offer one. The openai_compatible block does, so that is the route to a flat-rate endpoint.

This is not a workaround so much as the supported path for any third-party Claude endpoint. Set it up once and every assistant request in Zed draws from the same unlimited plan.

Because the connection is defined in one place, switching the whole editor onto unlimited is a single change: point api_url at the flat-rate gateway and every assistant feature follows. There is no per-feature billing toggle to hunt down, which keeps the unlimited setup simple. One api_url change is genuinely all that stands between a metered editor and an unmetered one. After that, every assistant feature simply inherits the flat rate.

Set api_url to the gateway

In settings.json add a provider under language_models.openai_compatible and set api_url to the gateway with /v1 on the end. Zed targets /chat/completions below that, hitting the gateway's OpenAI-compatible surface.

With the meter gone there is no reason to be cautious about how often the assistant fires. You can leave it wired into your everyday flow and let it answer as often as you like.

That freedom changes habits in practice: instead of saving the assistant for big questions, you reach for it on small ones too — renaming, quick explanations, one-line refactors. Those tiny calls add up fast on a meter, which is exactly why having them covered by a flat rate feels so different. Those small, frequent calls are exactly the ones a flat rate makes effortless to keep using. Over a working week they add up to far more value than the occasional big query.

List the models you want

Zed does not discover models automatically on a custom endpoint, so declare them in available_models. Add claude-sonnet-4-5 for everyday edits and claude-opus-4-8 for hard work, each with a name, a display_name, and a max_tokens.

On metered billing people often stick to the cheaper model to save money. On unlimited you can keep the strongest one selected as your default, because every model in the list costs the same flat rate.

Set max_tokens to the model's real capacity rather than a conservative number, since on a flat-rate plan there is no cost reason to clip responses short. A fuller max_tokens lets Claude return a complete refactor or explanation in one go instead of stopping partway. Sizing the limits to the model rather than the wallet is a habit unlimited lets you finally adopt. The result is fuller, less truncated answers in the editor.

Key in the keychain

The unlimited key does not belong in settings.json. Zed reads it from the OS keychain or an environment variable named for the provider, such as AIPRIME_API_KEY, so your synced config never carries the secret.

One key covers the whole plan. Enter it once when Zed prompts and the same flat-rate plan backs every assistant call you make for the period.

If you run Zed on more than one machine, each just needs the same unlimited key supplied locally through its keychain or environment; the shared settings file stays secret-free. The plan does not care which device a request comes from, so your laptop and desktop both draw from the one flat rate. The shared-config, local-key pattern scales cleanly from one machine to several with no extra cost. Each device draws from the same plan regardless of where you are working.

Or go native with ACP

If you prefer native Anthropic behaviour, run Claude Code as a Zed agent over the Agent Client Protocol. Claude Code carries its own ANTHROPIC_BASE_URL to the unlimited gateway, so Zed itself needs no Anthropic URL.

This pairs naturally with unlimited, because Claude Code's agentic loops move a lot of tokens. On a flat-rate plan you let those loops run to completion instead of cutting them short to manage spend.

It is the strongest argument for going the ACP route on this plan: a full agent that explores a codebase and edits across files is precisely the high-volume work that metered pricing makes you ration. With unlimited backing Claude Code, you point it at a hard task and let it work without watching a counter. For heavy agentic work the ACP route plus a flat rate is about as friction-free as Claude in an editor gets. You point Claude Code at a task and let it run to completion.

// Zed settings.json
{
  "language_models": {
    "openai_compatible": {
      "aiprime": {
        "api_url": "https://unlimited.aiprimetech.io/v1",
        "available_models": [
          { "name": "claude-sonnet-4-5",
            "display_name": "Claude Sonnet",
            "max_tokens": 200000 }
        ]
      }
    }
  }
}
// Key goes in the keychain / AIPRIME_API_KEY env var, NOT this file.

Frequently asked questions

Can I run Zed on an unlimited Claude plan?
Yes — add an openai_compatible provider with api_url {SITE}/v1, declare your Claude models, and the flat-rate plan covers every assistant call.

Why use the openai_compatible block?
Zed's native Anthropic provider has no api_url override. The openai_compatible block does, so it is how you reach a custom unlimited endpoint.

Why is unlimited good for Zed?
The assistant is used constantly for inline edits and questions, accruing calls all day. Flat-rate unlimited keeps that cost fixed.

Are there token limits?
No token limits during the plan, only fair-use rate limits so the gateway stays fast for everyone.

AI Prime Tech is an independent API gateway. It is not affiliated with, endorsed by, or a reseller of Anthropic. Claude and related model names are trademarks of their respective owners.