Skip to main content
Zed drives its Agent Panel from providers declared in settings.json. It discovers nothing, which makes it the most explicit of these harnesses to configure and the one where a wrong figure stays wrong quietly. Read coding harnesses first.

​
The configuration surface

The provider key is openai_compatible, and each provider under it is named by an id you choose. The base URL key is api_url:
settings.json
Zed does not append /v1. Its documentation’s own example value carries the version segment, so include it.

​
Declare the models yourself

Zed does not ask this API what it can call. There is no discovery step: the models the Agent Panel offers are exactly the entries in available_models, and name is the string sent as the model field, so a pinned id goes straight in. display_name is what you see in the picker, and max_tokens is the model’s context window. Take both figures from GET /v1/models, which publishes the context window and the maximum output for every model. An entry also accepts max_output_tokens, and a wrong max_tokens is not corrected by anything on our side: it changes how much conversation Zed sends before it trims, which shows up as a request the provider refuses for length rather than as a settings error. Adding a model means editing this list again, which is the trade for a harness that never surprises you with a model you did not choose.

​
Put the key in the environment

Zed’s documentation asks you directly not to put API keys in settings.json, which is the right instruction for a file people share. Two places are supported: the provider settings panel in the application, which stores the key in your system keychain, and an environment variable whose name is derived from the provider id you chose, upper-cased with _API_KEY appended. For the id in the example above that is:
The environment variable takes precedence over the keychain entry when it is set and not empty.

​
Leave chat_completions alone

Each declared model may carry a capabilities block, and one of its keys decides which endpoint Zed calls. chat_completions defaults to true, which is the endpoint this site documents. Setting it to false tells Zed to use the Responses endpoint instead, which this site publishes nothing about: no schema, no examples, no refusal list. Leave the default in place, so that what a request does is something you can look up here rather than something you have to discover. The same block is where tool calling is settled, and it defaults in your favour: tools is true unless you turn it off. So a model whose catalog entry says tools is yes needs nothing added here, which is not true of every harness. images and parallel_tool_calls default to false, so turn those on deliberately for a model whose catalog entry supports them.

​
The maximum-output setting is worth finding

A coding harness that sends no max_tokens holds the model’s whole published maximum output against your balance on every request, which is what makes the spend rate cap bind early. Zed’s model entry carries both a max_output_tokens key and a max_tokens_parameter capability, and we did not verify which of them results in a max_tokens on the wire. If the cap starts refusing requests during an agent session, that pair is the first place to look. See rate limits and spend controls.

​
What we did not verify

  • That a Zed session against this API succeeds. Nothing was run.
  • Which of max_output_tokens and max_tokens_parameter puts a max_tokens on the request, or what Zed sends when neither is set.
  • Whether Zed has any feature that calls an embeddings endpoint. Its current documentation names no embeddings model, no semantic index and no such setting, and the Agent Panel’s own description of searching a codebase names no embedding step. An older experimental semantic search did use embeddings. We found no statement that it was removed, so we report the absence rather than asserting it.
  • Whether a project-local .zed/settings.json can carry language_models at all. The documentation neither permits nor forbids it, so do not rely on it and do not put a key there while it is untested.
  • Whether the openai_compatible provider key has always been called that. The documentation for it has moved recently and we could not establish what, if anything, it was renamed from, so an older settings file may use a different key.