Start the server

  1. Install a local model and start Ollama if it is installed separately.
  2. Open Settings → Models and API → Local API and select Start. The default base URL is http://127.0.0.1:1234/v1. Set another port before starting if needed.
  3. Copy the key and send it in the Authorization: Bearer … header. Copy a new key after restarting the server.

The server listens on this computer and stops when codef quits. Each request contains its own list of text messages.

Find the model ID

Replace CODEF_KEY with your key. On Windows you can use curl.exe. 

curl -H "Authorization: Bearer CODEF_KEY" http://127.0.0.1:1234/v1/models

Copy the model’s exact id and use it in place of MODEL_ID.

Get a response

Save this UTF-8 request body as request.json.

{
  "model": "MODEL_ID",
  "messages": [
    {"role": "user", "content": "Write a short greeting."}
  ]
}

Send the request from the same directory.

curl -H "Authorization: Bearer CODEF_KEY" -H "Content-Type: application/json" --data-binary "@request.json" http://127.0.0.1:1234/v1/chat/completions

The response text is in choices[0].message.content. Add "stream": true and pass -N to curl for SSE streaming. Use "stream_options": {"include_usage": true} for token usage at the end of the stream.

Supported endpoints and parameters

  • GET /v1/models lists installed Ollama models. Read the identifier from data[].id.
  • POST /v1/chat/completions accepts model and messages. It supports text messages with the system, developer, user, and assistant roles.
  • temperature from 0 to 2 and top_p from 0 to 1 control sampling. stop accepts up to four sequences that end generation.

In a Chat Completions client, enter the base URL shown in codef and the local key. Use a single text completion. Agent tools, images, and embedding endpoints are unavailable through this server. See the Node.js and Python examples.

Request limits

A request can contain up to 128 text messages and use up to 256 KiB. Set max_completion_tokens or max_tokens to a value from 1 to 8192. The default is 2048. The server runs up to two generations at once.

Resolve request errors

  • 401. Copy the key from the running server.
  • 404 model_not_found. Check the model ID against GET /v1/models.
  • Connection failed. Check that the server is running and that the port matches its base URL.