Examples
Runnable scripts live in examples/.
Run them with uv run python examples/<name>.py. Most need a provider key
(OPENROUTER_API_KEY etc.); the cache benchmark and offline tests don't.
| Example | What it shows | Needs |
|---|---|---|
live_demo.py |
ask(), charts, and SQL on a cheap model |
API key |
demo.ipynb |
executed end-to-end notebook (all features) | API key |
streamlit_app.py |
upload a CSV, ask, see answer + chart inline | API key, [demo] |
eval_demo.py |
multi-model eval leaderboard (--suite hard\|large\|messy\|bespoke) |
OPENROUTER_API_KEY, [eval] |
eval_wtq.py |
WikiTableQuestions public benchmark | OPENROUTER_API_KEY, [eval] |
cache_benchmark.py |
code-replay cache: zero-token repeat questions | none (deterministic) |
mcp_demo.py |
use an external MCP server's tools through the harness | API key, [mcp] |
advanced_wiring.py |
explicit Harness wiring with connectors, planner, subagents |
ANTHROPIC_API_KEY |
inspect_run.py |
RunResult / JSONL run inspection |
API key |
quickstart.py |
minimal Agent example |
ANTHROPIC_API_KEY |
The Streamlit demo
A multi-model eval run
uv run python examples/eval_demo.py --suite messy
# prints a cost leaderboard and writes evals/results/messy_<ts>.{json,md}
See Evaluation for the suites and graders, and the committed results summary.