praxen

Quickstart — your first Praxen report in 5 minutes

This walks you from “Praxen is installed” to “I have a real report” against one of the bundled examples — finbot, a deliberately vulnerable invoice-processing agent from the OWASP Agentic AI CTF. No editing your own agent required.

If you haven’t installed yet, do Installation first (one command from the marketplace).

---

1. Get a copy of the example workspace

If you don’t have a local copy of the Praxen repository yet, clone it (or unzip a release):

git clone https://github.com/open-agent-ai-security/praxen.git
cd praxen

The pieces you’ll point Praxen at are already inside examples/finbot/:

examples/finbot/
  WORKER_REMIT.md          ← the policy doc you'll verify against
  finbot-analysis.html     ← the committed report (what your run should approximately produce)
  finbot-findings.json     ← the same content as canonical JSON

A real first scan would use your agent’s source plus a remit you wrote. We’re using the pre-staged ones so the first run has no moving parts.

2. Get the FinBot source

The example was developed against the CineFlow Productions finbot from the OWASP Agentic AI CTF (see examples/README.md for the full provenance). Clone the source so Praxen has a real workspace to read:

git clone https://github.com/OWASP-ASI/finbot-ctf-demo.git ../finbot-src

(Any directory will do — ../finbot-src keeps the clone outside the Praxen tree and works the same on macOS, Linux, and Windows.)

3. Run the analysis

From a Claude Code session in the praxen repo directory, ask the agent:

Please run the behavior-verifier skill against ../finbot-src.
Use the Worker Remit at examples/finbot/WORKER_REMIT.md. Write outputs
to ./reports/finbot-quickstart/.

That’s the whole prompt. Praxen will:

1. Read the Worker Remit
2. Sweep the workspace at ../finbot-src
3. Score the six RAISE categories
4. Audit every remit rule
5. Surface compound attack chains
6. Write three files to ./reports/finbot-quickstart/

The skill prints an interim overview to stdout while it works. When it finishes you’ll have:

./reports/finbot-quickstart/
  finbot-findings-YYYY-MM-DD.json     ← canonical record
  finbot-analysis-YYYY-MM-DD-HHMMSS.html  ← human-readable report
  finbot-analysis-YYYY-MM-DD-HHMMSS.txt   ← plain-text summary

4. Open the report

open ./reports/finbot-quickstart/finbot-analysis-*.html

(open is macOS; on Linux use xdg-open, on Windows the file works in any browser.)

What you should see, top to bottom:

• A red CRITICAL status badge — FinBot is deliberately broken
• An Agent Remit panel restating what the remit says, alongside an Agent Structure panel summarising what Praxen found in the workspace
• A Behavior Summary — the dominant pattern (typically something like “framework offers safe primitives, code uses none of them”)
• A Remit Coverage table listing every rule with status (Verified / Gap / Partial / Vague / ENP)
• A Findings Register — Critical first, then High, Medium, Low, Informational, each with evidence and a recommended action
• A RAISE Maturity Posture wrap-up — a 0–5 weighted score across six categories

You can compare your fresh report against the published FinBot example report (rendered on GitHub Pages). It won’t be byte-identical (LLM analyses have run-to-run variance) but the dominant Critical themes, the broad RAISE shape, and the remit-coverage counts should be close. See tests/README.md for what “close” actually means for this target.

5. Now try your own agent

The pattern is identical with a real target:

1. Write a Worker Remit for the agent
2. Point Praxen at whatever evidence you have — source, deployment files, behavioral logs, governance docs
3. Read the report; iterate on the remit and the agent as needed

See Usage for the full set of input shapes and the running-an-analysis details.

---

If something went wrong

See the Troubleshooting section in usage.md. The most common first-run snags:

• “behavior-verifier skill not found” — restart Claude Code or run /reload-plugins
• render.py errored at the end — the LLM produced a malformed findings JSON; re-run with more context window or a more focused workspace path
• Context window auto-compacted during the run — Praxen wrote a draft manifest in ./reports/<agent>-draft-<timestamp>.md; tell the agent to read it and finish from there. See Usage § Large workspaces and context sizing.

This site is open source. Improve this page.