Zed with OpenCheese

← API quickstartView as Markdown

Zed's agent panel reaches OpenCheese through an openai_compatible provider in Zed's settings.json, speaking the OpenAI Chat Completions API at https://api.opencheese.dev/v1.

01Open settings.json

Zed keeps every user preference in one file. Open it from the command palette with zed: open settings file, or edit it directly:

Merge, do not replace: the same file holds your theme, keymap, fonts and language servers.

02Add the provider

The dashboard's Configuration page, Zed tab, generates this block with every model your plan includes and the limits each one has. By hand, one model looks like this:

{
  "language_models": {
    "openai_compatible": {
      "OpenCheese": {
        "api_url": "https://api.opencheese.dev/v1",
        "available_models": [
          {
            "name": "MODEL_ID",
            "display_name": "Model name",
            "max_tokens": 1000000,
            "max_output_tokens": 131000,
            "max_completion_tokens": 126000,
            "reasoning_effort": "low",
            "capabilities": {
              "tools": true,
              "images": true,
              "parallel_tool_calls": false,
              "prompt_cache_key": true,
              "chat_completions": true,
              "interleaved_reasoning": false,
              "max_tokens_parameter": true
            }
          }
        ]
      }
    }
  }
}

• api_url keeps its /v1: Zed appends /chat/completions itself.
• If the file already has a language_models section, add only the OpenCheese key inside its openai_compatible object.
• max_tokens is the model's context window and max_output_tokens its longest answer: take them from context_window and max_output_tokens in /v1/models. Every model needs both, or Zed rejects the whole provider.
• reasoning_effort must be one of none, minimal, low, medium, high, xhigh or max; any other value makes Zed reject the whole provider. Leave it out for a model with no levels.
• Keep prompt_cache_key on: with it, each Zed thread appears as one session in the dashboard.
• images is true only for models whose modalities.input includes image.

03Give Zed the API key

The key never goes into settings.json: Zed has no field for it there. Either:

1. Open the agent panel, run agent: open settings, find OpenCheese under LLM Providers, paste the key into its API key field and press Enter. Zed keeps it in your system keychain.
2. Or set the environment variable Zed reads for this provider:

export OPEN_CHEESE_API_KEY='YOUR_OPENCHEESE_API_KEY'

Note the underscore: Zed builds the name from the provider ID and splits OpenCheese into OPEN_CHEESE. OPENCHEESE_API_KEY is never read. The variable also wins over the keychain, so unset it and restart Zed if you later switch keys.

04Pick an OpenCheese model (optional)

Zed keeps whatever model was selected before. To make an OpenCheese model the default for new threads, merge:

{
  "agent": {
    "default_model": {
      "provider": "OpenCheese",
      "model": "MODEL_ID"
    }
  }
}

provider is the plain provider ID, with no openai_compatible/ prefix. Restart Zed, or reload the window, after editing settings.json.

Troubleshooting

• No OpenCheese models at all: the JSON is malformed, or a model entry misses max_tokens or has an unknown reasoning_effort. Copy the generated block again rather than fixing it by hand.
• Models appear but every request fails: the key never reached Zed; repeat step 3.
• The model is not used: the provider ID must be OpenCheese in both blocks, spelled the same way.