examples

Examples

Runnable demonstrations of the eager-tools API.

File	What it shows	Requires
01_minimal.py	Eager dispatch end-to-end against an in-memory fake stream. No network.	nothing
02_anthropic_live.py	Real Claude stream + 3 fake tools. Watch tools fire as their JSON blocks seal.	ANTHROPIC_API_KEY, pip install anthropic
03_openai_live.py	Same harness against GPT-4o.	OPENAI_API_KEY, pip install openai
04_cancellation.py	Mid-flight stream.cancel() releases all in-flight tool tasks cleanly.	nothing
05_openrouter_live.py	Same harness against OpenRouter (OpenAI-compatible API, any tool-capable model).	OPENROUTER_API_KEY, pip install openai

Running

From a venv with the adapter packages installed:

python examples/01_minimal.py
ANTHROPIC_API_KEY=sk-ant-... python examples/02_anthropic_live.py
OPENAI_API_KEY=sk-...        python examples/03_openai_live.py
python examples/04_cancellation.py
OPENROUTER_API_KEY=sk-or-... python examples/05_openrouter_live.py

For development, you can use either adapter package's venv (each declares its own anthropic/openai extra under [project.optional-dependencies] dev):

cd packages/eager-tools-anthropic && uv sync --extra dev
uv run python ../../examples/01_minimal.py

What you should see

01_minimal and the live examples log a timeline like:

[   1.1 ms] stream opened — eager dispatch active
[ 102.0 ms] sealed fetch_weather   latency=100.8 ms — dispatched eagerly
[ 102.1 ms]   ↳ fetch_weather started
[ 203.4 ms] sealed fetch_stock     latency=101.3 ms — dispatched eagerly
...

The "sealed" lines arrive during the stream — well before message_complete. Total wall clock ≈ duration of the slowest tool, not the sum of all tools.