More screenshots
Works with the AI Agents you already use
MCP Config Audit — What Your MCP Config Gives Away
Reads your mcp.json and reports the unpinned package that fetches new code at every start, the token sitting in the file, the endpoint that is plain HTTP, and what runs without asking you.
Free
MCP Config Audit — What Your MCP Config Gives Away
Example session with this skill installed
python3 scripts/config_scan.py .mcp.json --baseline .mcp-baseline.json
mcp.json:
{"mcpServers": {"tickets": {"command": "npx",
"args": ["-y", "example-tickets-mcp"],
"env": {"TICKETS_API_TOKEN": "tk_live_9f2c..."}}}}
- Read your context and instructions
- Compiled the mcp config audit
mcp-config-audit .mcp.json (mcpServers)
servers 6 findings 13 high 10 medium 1 low 2
HIGH C1 unpinned package tickets example-tickets-mcp
HIGH C1 unpinned package search mcp-search-proxy@latest
HIGH C2 credential in the file tickets env.TICKETS_API_TOKEN
HIGH C2 credential in the file browser headers.Authorization
HIGH C3 plaintext transport browser http://mcp-browser.example/sse
HIGH C4 broad filesystem root files /
HIGH C5 auto-approved actions ops *
HIGH C6 shell execution ops bash -c curl ... | sh
MEDIUM C1 unpinned package db ghcr.io/example/mcp-postgres:latest
LOW C7 dangerous flag files -y
FAIL findings at or above 'high'
report mcp-config-report.md
What to do next
- Pin every package you kept: example-tickets-mcp@2.3.1, mcp-search-proxy==1.9.2, and a digest for the Postgres image.
- Replace the two literals with the client's variable syntax; keep the values in your secret store.
- Swap http:// for https:// on the browser endpoint, or run that server locally over stdio.
- Grant /Users/dev/project instead of / to the filesystem server.
- Remove the wildcard auto-approve; approve writes, sends and deletes in code.
- Re-run with --write-baseline once the diff is reviewed.
Connects securely to your tools. The creator never sees your data.
About this skill
Every MCP security write-up starts with the tool descriptions. This starts one step earlier, at the file that decides what runs at all.
An mcp.json is a launch script with credentials in it. npx -y some-server resolves a package at every start, so the code your agent runs can change without a commit. A token in env gets committed, synced and shared, and its reach becomes the server's reach. "autoApprove": ["*"] removes the one control that still works when a model is fooled. None of that appears in the tool list, so a tool-contract scan cannot see it — this reads the other half.
Point it at the config your client actually uses. It parses Claude Code, Claude Desktop and Cursor mcpServers blocks and VS Code's servers block, then reports each server's launch line, whether it is pinned, and every finding with the exact evidence and a fix written for your client's variable syntax.
- Finds unpinned
npx/uvx/bunxpackages,@latest, and:latestimages that fetch new code at startup - Flags literal tokens, bearer headers and database URLs sitting in the file instead of a variable reference
- Catches plain-HTTP endpoints,
/or$HOMEfilesystem roots, shell pipes, and sandbox-disabling flags - Reports auto-approved actions, especially wildcards and anything that writes, sends, deletes or pays
- Hashes each server so
--baselineshows a changed launch command or credential the next time you look - Writes a Markdown report and JSON summary, and fails CI at a severity threshold you choose
Runs on the Python standard library, needs no key and makes no network calls.
The shipped example catches 13 findings across 6 servers, including a support server launched unpinned with a live token and a shell server that auto-approves everything. The clean example reports zero, which is how you know it is not just printing warnings.
Good to know: this reads configuration only. It does not audit the server's own code or runtime behaviour, and config roots and allowlists are cooperative boundaries — a compromised server can still reach whatever its process can. TOML configs (Codex) are not parsed. It is a review aid for the launch line, not a safety verdict on the server.

Audits config files only, not server source code or runtime behavior. Does not support Codex TOML. Requires Python 3.10+.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 1 day ago
- Free with an Agensi account
Requires