How to Use TokenAir with Goose: Tool-Calling Limits
TokenAir can significantly lower Goose model spend when a lower-cost model handles suitable agent work without breaking tool calls. Goose supports custom OpenAI-compatible endpoints and makes tool calling central to its agent loop. Test that loop before regular use.
Compatibility verdict
Limited setup
Goose documents a custom OpenAI host and Chat Completions base path. TokenAir has not yet verified native tool calls and multi-turn tool results for each model.
A custom endpoint path exists, but important features or request shapes need separate testing.
Before you start
- A current Goose Desktop or CLI installation.
- A TokenAir API key and exact account-enabled model ID.
- A safe extension or built-in tool task with an obvious success condition.
Setup and compatibility steps
Step 1
Open provider configuration
In Goose Desktop, open Settings, Models, and Configure providers. In the CLI, run goose configure and choose Configure Providers.
Step 2
Choose the OpenAI provider
Select OpenAI, which Goose also uses for OpenAI-compatible endpoints. Enter the TokenAir API key and model ID.
Step 3
Split the host and request path
Use https://api.tokenair.ai as OPENAI_HOST and v1/chat/completions as OPENAI_BASE_PATH. Goose builds the full request URL from these two values.
Step 4
Run a tool-use smoke test
Ask Goose to perform one read-only tool action. Confirm the tool call, tool result, final answer, and total number of model turns.
Goose OpenAI provider values
Provider: OpenAI
OPENAI_HOST: https://api.tokenair.ai
OPENAI_BASE_PATH: v1/chat/completions
OPENAI_API_KEY: YOUR_TOKENAIR_API_KEY
Model: YOUR_TOKENAIR_MODEL_IDReplace placeholders with values from your own TokenAir account. Never commit an API key to source control.
How to verify it
- Goose validates and saves the OpenAI-compatible provider.
- A short message returns a complete response through TokenAir.
- A Goose tool call contains valid arguments and the result returns to the model.
- The agent reaches a useful final answer without an uncontrolled retry loop.
How to lower cost without hiding quality loss
TokenAir gives the client access to premium and lower-cost model choices through one OpenAI-compatible API. Savings depend on matching each task with a model that still passes review.
- Begin with a read-only task and inspect the number of agent turns before allowing mutations.
- Use a stronger model when tool-call reliability is worth more than a lower token price.
- Count extension calls, retries, and recovery prompts in cost per completed task.
Limits to know before production
- Goose relies on tool calling and warns that model choice affects agent performance.
- The host excludes /v1; the separate base path supplies v1/chat/completions for this provider flow.
- A compatible text response does not prove that Goose extensions and repeated tool results will work.
FAQ
Can Goose use a TokenAir OpenAI-compatible endpoint?
Yes for endpoint setup. Goose's OpenAI provider documents a custom host and supports OpenAI-compatible services.
Why is the Goose host missing /v1?
Goose asks for the origin and base path separately. Use https://api.tokenair.ai as the host and v1/chat/completions as the path.
Why is Goose still marked limited?
Goose depends on tool calling. TokenAir must verify the chosen model's tool call, tool result, and final-answer sequence before claiming full Agent compatibility.
Official sources and related TokenAir docs
Tool settings change between releases. We checked these notes on July 17, 2026. Review the linked tool docs again before a production rollout.
Try one workflow before changing the default.
Use the cost checkup to choose a bounded test, then compare cost per accepted result with your current route.