The tool: use_transformers

One @tool, the whole library. Three modes: discover, run, call. Nothing is hardcoded - the task list is read from transformers at runtime.

flowchart LR
    AG["🤖 Agent"] --> UT["🛠️ use_transformers"]
    UT --> DISC["🔍 discover"]
    UT --> RUN["▶️ run"]
    UT --> CALL["⚙️ call"]
    DISC --> REG["📚 transformers<br/><i>read at runtime</i>"]
    RUN --> REG
    CALL --> REG

    classDef a fill:#7C5CFF22,stroke:#7C5CFF,stroke-width:1.5px,color:#7C5CFF;
    classDef t fill:#22D3EE1f,stroke:#22D3EE,stroke-width:1.5px,color:#0F91A6;
    classDef r fill:#8B8B9414,stroke:#8B8B9466,stroke-width:1.5px,color:#6B6B76;
    class AG a; class UT,DISC,RUN,CALL t; class REG r;

Discover

Find the right task and model without memorizing the library.

use_transformers(action="tasks")        # every task + modality + default model
use_transformers(action="task_info", task="image-text-to-text")
use_transformers(action="inspect", target="pipeline")   # signature of anything

Run

High-level pipelines. inputs takes a path, URL, base64, text, dict, or array.

use_transformers(action="run", task="automatic-speech-recognition", inputs="clip.wav")

use_transformers(action="run", task="object-detection",
                 inputs="https://images.cocodataset.org/val2017/000000039769.jpg")

Vision tasks return structured data and save an image you can look at:

input	detection	depth

$ uv run vision.py
detections: ['remote', 'remote', 'couch', 'cat', 'cat']
depth map saved to: /tmp/strands_transformers/image_*.png

Call

For anything pipelines don't cover - VLA models, custom methods. Load components once, cache them, chain calls.

use_transformers(action="call", target="AutoProcessor.from_pretrained",
                 parameters={"pretrained_model_name_or_path": MODEL}, cache_key="proc")

use_transformers(action="call", target="cached:vla.predict_action",
                 parameters={"**": "cached:inputs", "unnorm_key": "bridge_orig"})

Cache shorthands

cached:key is a live object, cached:key.method calls a method on it, and "**" unpacks a cached mapping into kwargs - the idiomatic model.predict_action(**processor(prompt, image)).

---

Every call returns {status, content, data, artifacts} - data is always JSON-safe. Full signatures in the API reference.