Application recipes
On this page
All examples require an explicitly selected real provider, account and model. The desktop starts empty; there is no offline provider or canned response mode.
The real application runner uses a paired connection profile and records the actual response and reported usage. Its public prompts cover:
- Brandstorm: three brand directions for AgenticDriver, preserving its name, audience and product constraints.
- Literature review: experiments for evidence selection and ordering, grounded in notes from Lewis et al. (arXiv:2005.11401) and Liu et al. (arXiv:2307.03172).
- Email workspace: a draft of the actual alpha.5 release handoff, distinguishing published artifacts, pending PyPI approval and unconfirmed app adoption.
npm run test:real -- --connection /private/app/profile.json \
--provider YOUR_CONNECTED_INSTANCE --model YOUR_SELECTED_MODEL \
--case all --receipt /private/results.jsonThis makes three model requests and can consume the selected account's allowance.
Use --case brandstorm, literature or workspace for one request. The runner
checks the account-reported catalog and host permissions before inference. It never
switches providers, accounts, models or billing modes. Review the answers as well as
the completion receipt; a completed stream is not proof of answer quality.
For an embedded TypeScript runtime, use examples/brandstorm.ts,
examples/literature-review.ts or examples/email-workspace.ts. Set
AGENTICDRIVER_PROVIDER, AGENTICDRIVER_ACCOUNT_ID (the selected instance ID),
and AGENTICDRIVER_MODEL. Native sessions can additionally select
AGENTICDRIVER_BINARY and AGENTICDRIVER_ACCOUNT_DIRECTORY; API adapters require
their documented provider key on the execution host. The installed npm archive
contains equivalent .mts examples under examples/javascript.
The application-tool example performs an actual
text search over the Markdown file named by AGENTICDRIVER_INPUT_FILE, using a
real provider that supports application tools. It is a lexical search example;
semantic indexing requires a real embedding account and the retrieval API.
The approval example asks a real model for a brand
proposal and saves it in memory only after human approval.
Full app acceptance belongs to Brandstorm, LitAgent and AI Workspace. These SDK examples do not send mail, modify a research library or save product designs. See real connection setup and evidence.