ContextGuard

Check what your AI agent reads before it trusts it.

Agents that keep notes about themselves fill those notes with duplicates, rules that contradict each other, old facts, pasted secrets, and sometimes instructions someone else slipped in. ContextGuard audits those files with an offline script, then helps you decide what to change.

What it finds

• Secrets (provider keys, private keys, tokens, passwords in URLs); values are masked in the report
• Injected instructions ("ignore previous instructions", hidden HTML comments, chat-control markup)
• Invisible or direction-changing characters
• Risky commands (download-and-run, recursive delete, force push, TLS off, standing permission to act without asking)
• Contradictions between rules, and the same setting with two values
• Duplicates and near-duplicates
• Stale, expired or undated TODO entries
• References to files that do not exist
• Instructions to send data to outside addresses
• Size and bloat

What you get

1. A verdict line and a findings table with severity and line numbers
2. Cleaned copies (secrets redacted, invisible characters removed, exact duplicates removed) plus a change log; your originals are never edited
3. A proposed edit for each decision that is yours to make
4. A list of secrets to rotate
5. A "not measured" list and a "My reading, not checked by the script" section

Included: a deliberately flawed sample AGENTS.md and LEARNINGS.md with the expected report, so you can see the output before using your own files.

How to use it

1. Unzip the package into your agent's skills or working folder
2. Give your agent the files to check (up to 20 text or Markdown files)
3. Optionally say which hosts you trust and how old an entry may be before it counts as stale (default 180 days)

Where it runs and what you need

ContextGuard is a download package. You unzip it and use it in your own environment. It does not run through Capafy's cloud, and no AI model is included in the price.

You provide:

• An AI agent or assistant that can read a SKILL.md file, read and write files in a working folder, and run shell commands. The instructions are written for a Claude-style agent; other agents have not been tested.
• Python 3 (standard library only, nothing to install; developed on Python 3.13, other 3.x versions not tested).
• Your own model or provider account. Any model usage fees are charged by your provider, not by this package.

Data