Hard-coded secrets
API keys or tokens sitting in an MCP server's config, where every agent session can read them.
Open source · Free · Runs locally
The AI coding-assistant scanner is a free, open-source tool with 14 checks for the MCP servers, Agent Skills and memory files your AI coding agent loads automatically; it runs on your own machine and only reads files. Every finding is explained in plain English.
A free, open-source scanner for everything your AI coding agent loads — with every finding explained in plain English, and nothing leaving your machine.
14
deterministic rules, the same result every time
0
network calls; nothing leaves your machine
3
kinds of file it reads: MCP servers, Agent Skills, memory files
100%
precision and recall on a generated benchmark corpus: self-graded, not yet independent validation
If you use Claude Code, Claude Desktop, or Cursor, your agent already trusts a growing pile of things without asking you each time: MCP servers (outside tools it can call), Agent Skills (packaged instructions it can follow), and memory files like CLAUDE.md that it reads on every session. Most teams have never seen the full list in one place — and security research keeps finding real, exploitable problems hiding in exactly these things: servers with no authentication, skills bundling malicious scripts, instruction files quietly poisoned to make an agent misbehave persistently.
The scanner inventories all of it on your own machine and tells you exactly what needs attention — in language a manager with no security background can act on, not just a security engineer.
You need Node.js 20 or later and git. Clone the repository and build it once; nothing is installed globally and no account is created.
git clone https://github.com/anthonyhammock/BlissPoint-Security.git
cd BlissPoint-Security
npm ci
npm run buildThe npm package is coming; until it is published, the scanner runs from source like this.
From that folder, give it the path of the project you want checked: the one that holds your agent's configuration, with a .claude/ or .cursor/ directory, a CLAUDE.md, or an MCP server config.
node dist/cli.js scan --dir ~/code/my-appIt also lists what your agent loads from your own settings on this machine.
The scanner opens config files and text files and looks at their contents. It never executes, imports or runs anything it finds, never changes a file, and never sends anything off your machine — there is no network call and nothing is recorded anywhere. Run it as often as you like.
First an inventory — how many MCP servers, skills and memory files it found — then the findings, most severe first. Each one names its rule, the file and line, what this means, and what to do, and maps to a published OWASP category. A clean run says "no known issues found by the current rules" — deliberately not "secure".
node dist/cli.js lock records a baseline; scan --verify then flags anything new, changed or removed since. scan --sarif results.sarif writes results for GitHub's code-scanning UI, and the ready-made GitHub Action in the repository builds the scanner from source and runs it on every push.

14 deterministic rules — no LLM involved, no network call, the same result every time on the same input. Every finding maps to the OWASP MCP Top 10 or the OWASP Agentic Skills Top 10, with a plain-English explanation and a concrete fix — never a bare "vulnerability detected". Among them:
API keys or tokens sitting in an MCP server's config, where every agent session can read them.
An MCP server launched without a pinned version — the classic "rug pull" setup where tomorrow's update isn't today's code.
A remote MCP server reachable with no visible authentication at all.
Shell commands assembled from untrusted input inside a server or skill.
Invisible characters in a skill or memory file — the standard way to smuggle instructions past a human reader.
Text telling the agent to send, upload or forward your data to an outside address.
An instruction to download a script and run it straight away, sight unseen.
Deterministic rules, a benchmark you can run yourself, and findings you can act on in minutes.
Every rule is scored against a generated benchmark corpus that ships in the repository: 56 malicious and 101 clean samples, each written by a script to exercise the rules. It scores 100% precision and recall there, but the project wrote both the corpus and the rules, so that is self-graded, not independent validation. Run it yourself with npm test.
A finding names its rule, the file and the line, says what it means in plain English, and tells you what to do — ranked most severe first and mapped to the OWASP category it belongs to. Five minutes per finding is the design target, and most take less.
"No known issues found by the current rules" is a precise statement: these 14 rules, this inventory, this moment. Lock a baseline and every later change to what your agent loads is flagged the instant it appears. The rule set and the tested platforms are listed in the project README.
See an AI assistant built with these safeguards: Aiva at BlissPoint Access asks before anything goes out, never sees passwords, and every action can be undone.
Anyone whose team uses an AI coding agent: developers, the person who looks after a small company's laptops, and managers who need to know what these tools can reach. Run it before you install a new MCP server or skill, after you copy a config from a blog post or a teammate, and on a schedule in your build pipeline so a changed file is caught the day it appears. If you hire contractors who bring their own agent setup, ask them to run it on the shared project and send you the summary.
It is one layer, not the whole job. Pair it with the basics on the Security hub: multi-factor login, same-day removal of leavers' access, and private remote access through a WireGuard VPN on your own server.
| Check | This scanner | Typical secrets or dependency scanner |
|---|---|---|
| Hard-coded API keys in MCP server configs | Yes | Often, if the file is in the repo |
| Unpinned or unauthenticated MCP servers | Yes | No |
| Hidden Unicode and injected instructions in skills and memory files | Yes | No |
| Known vulnerable packages in your app | No | Yes |
| Baseline lock to flag new or changed agent tools | Yes | No |
Keep your existing tools; this one covers the part of the stack they don't read.
A tool that lists everything your AI coding agent loads on its own, such as MCP servers, Agent Skills and memory files like CLAUDE.md, and flags risky settings and hidden instructions in them, with a plain-English fix for each.
No. It runs on your own machine, only reads files, and makes no network call. Nothing is uploaded, logged or shared.
Claude Code, Claude Desktop and Cursor configurations, plus MCP server configs, Agent Skills folders and memory files in any project you point it at. The README lists the tested platforms.
Yes. It is open source under the MIT license, with no account, no usage limit and no paid tier.
No, and the scanner says so. A clean run means no known issues were found by its current rules. It doesn't replace reviewing what you install or a full security review.
Those look at your application code and packages. This scanner looks at what your AI agent trusts: the tools it can call, the skills it follows and the instructions it reads every session, which other scanners usually skip.
Run it on every project, as often as you like, with nothing to sign up for. Security disclosure policy · questions to solutions@blisspointanalytics.com.