Claude Code MVP Quickstart

This is the shortest product-facing path through Ardur today from a source checkout. It is meant for the current pre-release dev branch and source installs; it is not a tagged package-manager release path yet.

Use it in two modes:

• No-key confidence check: no Anthropic/OpenAI token required. It verifies source/local-wheel install, ARDUR.md profile creation, ardur protect claude-code, ardur doctor-claude-code, a simulated Claude Code hook allow and deny path, and ardur claude-code-report chain verification.
• Live Claude Code demo: requires the local claude binary to already be installed and authenticated. Ardur does not perform login, account changes, or provider setup.

Claim boundary

1. Install from source

From a fresh checkout of this branch:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e python/
ardur --help

Keep the virtualenv active for the rest of the walkthrough so Claude Code hooks can find the same installed ardur package.

2. Run the no-key evidence harness

This does not call a live LLM provider. It uses temporary HOME, project, Ardur home, and evidence directories, then writes a redacted shareable bundle.

python3 scripts/run-rwt-phase1-fresh-user.py \
  --expected-origin-dev "$(git rev-parse --short=12 origin/dev)" \
  --output-dir /tmp/ardur-rwt-phase1

python3 -m json.tool /tmp/ardur-rwt-phase1/bundle.redacted.json | less

Expected result for a clean source checkout:

• bundle status is PASS
• RWT-1 is PASS for install/profile/protect/doctor
• RWT-2 is PASS for actual hook CLI fixture allow/deny receipts
• RWT-3 is PASS, SKIP_GATED, or SKIP_UNSUPPORTED depending on whether a logged-in claude binary is available; a skip is the honest no-key result, not a hidden failure
• secret_scan_hits is 0
• raw_secret_values_copied is false

3. Run a live Claude Code session

Only run this if claude is already installed and logged in. The demo creates a temporary project and a local .vibap home under that project.

ARDUR_REPO="$(pwd)"
DEMO_PROJECT="$(mktemp -d "${TMPDIR:-/tmp}/ardur-claude-demo.XXXXXX")"
cd "$DEMO_PROJECT"

printf 'alpha\nbeta\ngamma\n' > notes.txt
export VIBAP_HOME="$DEMO_PROJECT/.vibap"

ardur profile init --template read-only --path ARDUR.md
ardur protect claude-code \
  --profile ARDUR.md \
  --home "$VIBAP_HOME" \
  --plugin-dir "$ARDUR_REPO/plugins/claude-code"
ardur doctor-claude-code \
  --home "$VIBAP_HOME" \
  --plugin-dir "$ARDUR_REPO/plugins/claude-code"

VIBAP_HOME="$VIBAP_HOME" claude \
  --plugin-dir "$ARDUR_REPO/plugins/claude-code" \
  -p "Use Read to summarize notes.txt. Do not edit files or run shell commands."

ardur claude-code-report --home "$VIBAP_HOME"

The report should find the receipt chain under $VIBAP_HOME/claude-code-hook/<trace_id>/receipts.jsonl, verify signatures and chain links, and summarize compliant, violation, and unknown outcomes. If the model attempts Bash, Edit, or Write, the read-only profile should return a Claude Code deny decision and still preserve the signed violation receipt.

4. Read the result correctly

Ardur evidence is strongest at the local tool boundary. Treat the report as a verified statement about what Claude Code exposed to local hooks and what Ardur allowed, denied, or could not see. Do not use it to claim provider-internal reasoning visibility, complete shell side-effect capture, production eBPF coverage, or package-manager release readiness.

Related references:

• plugins/claude-code/README.md
• docs/reference/cli.md
• docs/reference/ardur-md-profile.md
• docs/coverage-map.md