Zed with a custom LLM provider

macOS and Linux

~/.config/zed/settings.json

{
  "language_models": {
    "anthropic_compatible": {
      "aigrow": {
        "api_url": "https://aigrow.gg",
        "available_models": [
          {
            "name": "claude-opus-5.5",
            "display_name": "Claude Opus 5.5",
            "max_tokens": 1000000,
            "capabilities": {
              "tools": true,
              "images": false,
              "prompt_caching": true
            }
          },
          {
            "name": "claude-sonnet-5.5",
            "display_name": "Claude Sonnet 5.5",
            "max_tokens": 1000000,
            "capabilities": {
              "tools": true,
              "images": false,
              "prompt_caching": true
            }
          }
        ]
      }
    },
    "openai_compatible": {
      "aigrow-openai": {
        "api_url": "https://aigrow.gg/v1",
        "available_models": [
          {
            "name": "gpt-6-astra",
            "display_name": "GPT-6 Astra",
            "max_tokens": 1050000,
            "capabilities": {
              "tools": true,
              "images": false,
              "chat_completions": true
            }
          },
          {
            "name": "grok-4.7",
            "display_name": "Grok 4.7",
            "max_tokens": 500000,
            "capabilities": {
              "tools": true,
              "images": false,
              "chat_completions": true
            }
          }
        ]
      }
    }
  }
}

404 on the Claude provider

The api_url ends in /v1. On the Anthropic compatible provider it is https://aigrow.gg.

404 on a GPT model

capabilities.chat_completions is false, which sends the model to the Responses API. Set it to true.

My tool says 401 or authentication error

The key is missing or mistyped. Copy it again from API access on the dashboard, and make sure it is sent as the key, not as a model name. A revoked key fails the same way: create a new one and swap it in.

I get 402 or a message about credit

Your balance is empty. Open API access on the dashboard, redeem a code or buy another pack, and the next request goes through. Nothing else needs to change.

It says 404 or Unknown model

The model has to be one the endpoint serves, for example claude-opus-5.5. The provider’s own spellings work too (claude-opus-5-5, anthropic/claude-opus-5.5), but a dated id or any model outside the catalogue is rejected before the request is forwarded, and nothing is billed. Request GET /v1/models with your key for the live list, or read the ids on the price list.

Requests are rejected with 429

The endpoint’s capacity is shared by every buyer: about 20 requests a minute and 4 at once. Each 429 carries a retry-after header, and the OpenAI and Anthropic SDKs wait that long and retry on their own. In a loop of your own, wait the seconds it names or lower your concurrency.

The model is unavailable right now

A 404 that says the model is not available right now means it is briefly down at the upstream provider, not a problem with your key or balance. Retry in a moment or send the request to another model.

I get a 502 or 503

A transient upstream or configuration hiccup, not a charge against your balance. Retry the request; if it keeps failing, reach us through the store and we will look into it.

A developer working all day, with most of each request read back from the prompt cache, priced at the list rate and through credit. Change the model to compare.

Anthropic compatible or OpenAI compatible in Zed?

Both work. Anthropic compatible, with api_url https://aigrow.gg, suits Claude models and prompt caching. OpenAI compatible, with api_url https://aigrow.gg/v1, runs GPT and Grok.

Where does Zed keep the key?

In the system keychain when you paste it in the provider settings, or in an environment variable named after the provider, such as AIGROW_API_KEY. Zed’s docs say never to put it in settings.json.

Should I turn on prompt caching?

Yes, for Claude models on the Anthropic compatible provider. Zed then marks the parts of the prompt to cache, and cache reads cost less than fresh input.

Sources: Zed Use API access · Zed LLM providers · Zed Configuring Zed. Read 21 October 2026.