Ardur Examples

Working examples of Ardur governing AI agents across major frameworks and local assistant surfaces. Some directories are runnable today; deferred directories are marked as adapter specs, not shipped capability.

Status

The runnable framework directories (langchain-quickstart/, langgraph-quickstart/, autogen-quickstart/) ship a demo.py entrypoint and, where applicable, a Dockerfile that produces the published rahulnutakki/ardur-demo:* images. They share helpers under _shared/ — provider selection, SVID fetch, Biscuit issuance, governed-session setup, receipt-chain verification, end-of-session attestation. No model identifiers are hard-coded in any of these files; provider config is sourced from environment variables at runtime (see CONTRIBUTING.md “No specific LLM model names” rule).

The deferred adapter directories carry READMEs that describe the dependency footprint and file layout the next import wave will produce. They are not advertised as runnable examples until code and tests land.

Running the mission examples (today, no agent required)

cd ../python
pip install -e .

# Issue and verify a passport. ardur issue takes mission claims via flags,
# not a JSON file — the example mission files under missions/ are reference
# documents for the spec layer. To exercise the protocol path:
ardur issue \
  --agent-id alice \
  --mission "summarize sales from sales/q1.csv into reports/" \
  --allowed-tools read_file write_report \
  --resource-scope 'sales/*' 'reports/*'

ardur verify --token <token-from-issue-output>

That exercises the core protocol surface end-to-end — mission compilation, passport issuance, signature, verification — without an LLM or framework in the loop. It’s the fastest way to confirm a local install actually works.

Why deferred adapters instead of one big drop

Each framework has its own tool-call interface, its own session-state model, and its own integration point where Ardur’s governance proxy attaches. LangChain tool callbacks look nothing like AutoGen’s FunctionTool registration; LangGraph’s state graph wants the verifier wrapped around node transitions; the coding-agent CLI integration wires in via a hook lifecycle, not a Python import. Lifting these as one monolithic commit would conflate unrelated breakage. Per-framework directories let each adapter land, get reviewed, and run CI on its own.

CI for examples

The current CI surface is the repo-wide Python and Go workflow in .github/workflows/tests.yml, plus CodeQL, link-check, secret-scan, format validation, and the Hugo site build. The framework quickstarts are runnable from the checked-in example directories, but there is not yet a dedicated examples-smoke.yml workflow for every adapter. Treat that as future hardening, not current gate coverage.