Commit bbfd2f2

model: str,

model: str,

model: ModelComposition | str,

model: str,

model: str,

model: str,

model: str,

model: str,

model: str,

model: str,

You can configure the model provider by setting the `ASKUI__VA__MODEL_PROVIDER` environment variable, e.g., `"bedrock"` to use Bedrock models. All the models you pass via `model` parameter of `act()`, `get()`, `locate()` will be prefixed with `"bedrock/"` in this case, e.g., if you call `agent.act("do something", model="anthropic.claude-sonnet-4-20250514-v1:0")`, it will be called as `agent.act("do something", model="bedrock/anthropic.claude-sonnet-4-20250514-v1:0")` under the hood. Alternatively, just prefix the model name(s) you pass via `model` parameter, e.g., `agent.act("do something", model="bedrock/anthropic.claude-sonnet-4-20250514-v1:0")` or `agent.act("do something", model="vertex/claude-sonnet-4@20250514")`.

At the time of writing, the following model providers are available:

- `"bedrock"`: Use models hosted on AWS Bedrock.

- `"vertex"`: Use models hosted on Google Vertex AI.

- `"anthropic"`: Use models hosted behind Anthropic API.

- `"askui"`: Use models hosted behind AskUI API.

**IMPORTANT:** If you pass a `model` argument at construction, this is not going to be prefixed with the model provider, e.g., if you call `VisionAgent(model="claude-sonnet-4-20250514", model_provider="askui")`, and later call `agent.act("do something")`, it will be called as `agent.act("do something", model="claude-sonnet-4-20250514")` and not as `agent.act("do something", model="askui/claude-sonnet-4-20250514")` under the hood. If you want to use a provider per default, just prefix the model name(s) you pass via `model` parameter, e.g., `VisionAgent(model="askui/claude-sonnet-4-20250514")` or `VisionAgent(model={"act": "askui/claude-sonnet-4-20250514"})`.

You can also set the `model` parameter of an agent via environment variable, e.g., `ASKUI__VA__MODEL`. For complex values, just use json, e.g., `ASKUI__VA__MODEL={"act":"askui/claude-sonnet-4-20250514"}`.

**IMPORTANT:** Keep in mind that the model name may differ between providers and not all providers may support a model (see https://docs.claude.com/en/docs/about-claude/models/overview). `askui` uses the same model names as `anthropic`. For `vertex` see https://docs.claude.com/en/api/claude-on-vertex-ai and for `bedrock` see https://docs.claude.com/en/api/claude-on-amazon-bedrock.

**IMPORTANT:** Keep in mind that when using a custom model that you may have to pass different `settings` to act as the settings support, e.g., tool or betas, differs between models, e.g., `agent.act("do something", model="askui/<a-special-model>", settings=ActSettings(tools=[ASpecialTool()]))`.

The following environment variables control authentication and behavior per provider. Variables marked as required must be set for that provider, unless your environment provides credentials through instance/role bindings or local SDK configuration.

- `ASKUI__VA__MODEL_PROVIDER` (str, optional): Provider prefix to apply automatically (e.g., `bedrock`, `vertex`, `anthropic`, `askui`). Per default, no provider prefix is applied, e.g., `claude-sonnet-4-20250514` is going to be called as `claude-sonnet-4-20250514` and not as `bedrock/claude-sonnet-4-20250514` under the hood.

- `ASKUI__VA__MODEL` (str | json): Default model or per-capability map. Example: `{"act":"bedrock/claude-sonnet-4-20250514"}`.

- `ASKUI_WORKSPACE_ID` (UUID, required): Workspace to route requests to.

- `ASKUI_TOKEN` (str) or `ASKUI__AUTHORIZATION` (str): Exactly one required. If `ASKUI__AUTHORIZATION` is set, it is used verbatim as the `Authorization` header. Takes precedence over `ASKUI_TOKEN`.

- `ASKUI_INFERENCE_ENDPOINT` (url, optional): Override base endpoint. Default: `https://inference.askui.com`.

- `ANTHROPIC_API_KEY` (str, required): Anthropic API key.

- `ANTHROPIC_AUTH_TOKEN` (str, optional): Anthropic auth token.

- `ANTHROPIC_BASE_URL` (url, optional): Base URL to use for Anthropic API.

- Uses standard AWS credential resolution. Set one of:

- `AWS_PROFILE` (str) or static credentials `AWS_ACCESS_KEY_ID` (str), `AWS_SECRET_ACCESS_KEY` (str), optional `AWS_SESSION_TOKEN` (str).

- Region (required): `AWS_REGION` (str) or `AWS_DEFAULT_REGION` (str).

- Any other AWS SDK configuration is respected (env, shared config/credentials files, instance role, SSO, etc.).

- `ANTHROPIC_BEDROCK_BASE_URL` (url, optional): Base URL to use for Bedrock API.

- Uses Google Application Default Credentials (ADC). Common setups:

- `GOOGLE_APPLICATION_CREDENTIALS` (path): Service account JSON key file.

- Or gcloud-authenticated user with `gcloud auth application-default login`.

- Project and location are resolved from ADC and/or environment; typical envs if needed in your setup: `GOOGLE_CLOUD_PROJECT` (str), `GOOGLE_CLOUD_LOCATION` (str). Consult your org’s Vertex configuration if these are required.

- `CLOUD_ML_REGION` (str): Region to use for Vertex AI.

- `ANTHROPIC_VERTEX_BASE_URL` (url): Base URL to use for Vertex AI.

- `ANTHROPIC_VERTEX_PROJECT_ID` (str): Project ID to use for Vertex AI.

import os

from askui import VisionAgent

os.environ["ANTHROPIC_VERTEX_PROJECT_ID"] = "test-project"

os.environ["CLOUD_ML_REGION"] = "europe-west1"

with VisionAgent() as agent:

agent.act("do something", model="vertex/claude-sonnet-4@20250514")

with VisionAgent(model_provider="vertex") as agent:

agent.act("do something", model="claude-sonnet-4@20250514")

with VisionAgent(model="vertex/claude-sonnet-4@20250514") as agent:

agent.act("do something")

os.environ["ASKUI__VA__MODEL"] = '{"act":"vertex/claude-sonnet-4@20250514"}'

with VisionAgent() as agent:

agent.act("do something")

If you would like to configure you own model provider, e.g., let's say `"openai"`, just use provider string as key in the model registry as described in [Your own custom models](#your-own-custom-models) section but suffix it with a `"/"`, e.g., `"openai/"`, in order to differentiate it from regular model names. If you then, later call a model of that provider, e.g.,

`agent.act("do something", model="openai/gpt-4o")`, the request will be routed to the model (api client) implementation that is the value of `"openai"` key in the model registry and the model passed to the underlying model (api client) implementation will be `"gpt-4o"`.

**IMPORTANT:** If you configure both a provider as well as a model prefixed with the provider prefix in the model registry, e.g., both `"openai/"` and `"openai/gpt-4o"`, the model has is going to be used and not the provider, e.g., `agent.act("do something", model="openai/gpt-4o")` is going to use `"openai/gpt-4o"` and not `"openai/"`, and `"openai/gpt-4o"` is going to be passed to the model (api client) implementation as `model` parameter instead of `"gpt-4o"`.