Providers¶

Provider	CLI	Extra	Credentials
Anthropic	`anthropic:`	`anthropic`	`ANTHROPIC_API_KEY`
OpenAI	`openai:`	`openai`	`OPENAI_API_KEY`
Azure OpenAI and Foundry	`azure:`	`azure`	`AZURE_OPENAI_ENDPOINT`, and `AZURE_OPENAI_API_KEY` or Entra ID
Amazon Bedrock	`bedrock:`	`aws`	the AWS credential chain
Google Gemini and Vertex AI	`google:`	`google`	`GEMINI_API_KEY`, or ADC with `GOOGLE_GENAI_USE_ENTERPRISE=true`
Ollama	`ollama:`	`ollama`	none
Your own function	`CallableTarget`	—	—

modelscape run pressure --target anthropic:<model> --target ollama:qwen3:14b

Credentials come from the environment. To keep them in a file, copy .env.example to .env and run with uv run --env-file .env modelscape run ....

In Python, every target takes model, max_tokens, and options for extra request fields; all but Anthropic also take temperature and top_p. Every setting that affects behavior is recorded with the results; credentials never are.

Targets hold open connections, except Bedrock and CallableTarget. When you run them yourself, await target.aclose() in the same event loop once the run is over; the CLI does this for you.

Azure: model is the deployment name.
Bedrock: model is a model ID or inference profile; pass region.
Ollama: pass think=False to stop a reasoning model from thinking, and set version to the model's digest, since a tag can point to new weights.

Local models¶

Ollama runs open models on your own hardware, so no data leaves the machine. Any other server with the OpenAI Responses API works too: set OPENAI_BASE_URL.

Live tests¶

Set a model, such as MODELSCAPE_OPENAI_MODEL, with its credentials, and run uv run pytest -m integration. Each configured provider answers one prompt and runs a small spec through modelscape run, check, and view.