Skip to main content
goose loads extra providers from JSON files in its custom_providers directory, so NagaAI needs no plugin and no goose release. The file on this page adds NagaAI as the nagaai provider: goose sends its requests to the NagaAI Responses API, streamed, and fills the model picker from /v1/models. The setup follows the goose docs on configuring a custom provider.

Before you start

You need a NagaAI API key and goose Desktop or the goose command-line tool installed.

Connect goose

1

Save the key

Export the key in the shell that starts goose, or add the line to your shell profile:
The file names this variable in api_key_env, so the key itself stays out of the file.
2

Download the provider file

On macOS and Linux:
On Windows, in PowerShell:
The file sets engine to openai, reads the key from NAGA_API_KEY, and lists five models with their context windows: anthropic/claude-sonnet-5.5, anthropic/claude-opus-5.5, openai/gpt-6.1-sol, deepseek/deepseek-v4-pro and minimax/minimax-m3.1-flash. With dynamic_models set to true, goose also loads every other model from the catalog.Its base_url is https://api.naga.ac/v1 and its base_path is v1/responses, so goose sends every request to POST /v1/responses, and goose reads the model list from GET /v1/models on the same host. goose sends the x-naga-supported-endpoint header with every request: on /v1/models it limits the list to the models the Responses endpoint serves, and the other endpoints ignore it.
3

Pick the provider

Restart goose and choose NagaAI in the provider list. In goose Desktop, open Settings > Models > Configure providers. In a terminal, run goose configure and choose Configure Providers. To keep the key in the goose keychain instead of the environment, enter it in the provider’s settings in goose Desktop.
4

Check the connection

Send one prompt:
The answer comes from anthropic/claude-sonnet-5.5 on NagaAI.

Context windows and costs

goose takes a model’s context window from context_limit in models. For a model that isn’t in the file, goose looks the id up in its own model catalog and, if it doesn’t find it, uses a 128,000-token window. To give another model its full window, add an entry with the value from limits.context_window of the model’s chat operation in /v1/models. The file carries no prices, so the costs goose shows come from its own catalog and are estimates. What each request cost is in Logs in your dashboard.

Troubleshooting

NagaAI is missing from the provider list

goose didn’t load the file. Check that it’s named nagaai.json, sits in the custom_providers directory for your system, and is valid JSON. goose also skips the provider when NAGA_API_KEY isn’t set. Set the key and restart goose.

The model picker shows only five models

goose couldn’t read /v1/models and fell back to the models in the file. Check that the file holds "base_url": "https://api.naga.ac/v1" and "base_path": "v1/responses", then restart goose.

401 invalid_api_key

goose sent no key or a wrong one. Check that NAGA_API_KEY holds an API key from the dashboard.

400 model_not_found

No model has this id, or the model doesn’t serve the Responses API. Pick a model from the list goose shows, or copy an id from GET /v1/models with the x-naga-supported-endpoint: responses header.