Zed: openai_compatible provider in settings.json

The Zed editor has an "OpenAI API Compatible" provider type built for exactly this. This page was checked against the docs for version 1.20.

Adding the provider

From the UI: Settings → AI → LLM Providers → Add Provider, which asks for a name, API URL, model id and context size. Or edit settings.json directly:

{
  "language_models": {
    "openai_compatible": {
      "uttapen": {
        "api_url": "https://api.uttapen.ir/v1",
        "available_models": [
          {
            "name": "anthropic/claude-sonnet-5",
            "display_name": "Claude Sonnet 5",
            "max_tokens": 200000,
            "max_output_tokens": 8192,
            "capabilities": { "tools": true, "images": true }
          },
          {
            "name": "openai/gpt-5-nano",
            "display_name": "GPT-5 nano",
            "max_tokens": 400000,
            "max_output_tokens": 1024
          }
        ]
      }
    }
  }
}

• name is the id we receive; a / in it is fine.
• max_tokens is the context window, not the output ceiling. That is Zed's naming, and it's where everyone trips up first.
• Always set max_output_tokens. Without it Zed sends no output ceiling at all, so our hold falls back to a default estimate; if the model writes more, the real amount is still deducted in full and the balance can dip negative (pricing).
• capabilities.tools is on by default, images is off. Turn images on for a model that really accepts them (vision).

The API key

Don't put the key in settings.json. Either enter it in Zed's provider screen, where it goes to the system keychain, or set an environment variable named after the provider id in upper snake case plus _API_KEY. For the id uttapen:

export UTTAPEN_API_KEY="sk-up-..."

Models and roles

Take ids from the model list: anthropic/claude-sonnet-5, openai/gpt-5, qwen/qwen3-coder or deepseek/deepseek-v3.2 for coding.

Zed also picks models for side jobs, all under the agent key as a provider-and-model pair:

{
  "agent": {
    "default_model": { "provider": "uttapen", "model": "anthropic/claude-sonnet-5" },
    "commit_message_model": { "provider": "uttapen", "model": "openai/gpt-5-nano" },
    "thread_summary_model": { "provider": "uttapen", "model": "openai/gpt-5-nano" }
  }
}

inline_assistant_model, compaction_model and subagent_model follow the same shape and fall back to default_model when unset. Commit messages and thread summaries are tiny jobs; sending them to something cheap like openai/gpt-5-nano keeps background spend near zero.

Edit prediction (inline completion) is configured separately and uses the /v1/completions path. If you enable it, keep its max_output_tokens small — it fires as you type.

Checking that it works

Open the Agent panel, select the uttapen model, ask something short, then find the request in Dashboard → Usage.

Troubleshooting

When the available balance can't cover the whole output ceiling, uttapen lowers it to fit (floor 1,024 tokens) and returns the granted value in the X-Uttapen-Max-Tokens header. 402 is only for the case where even that floor is unaffordable.

Other tools are in integrations; error codes are in errors.