What it does
Coding agents read instruction files (AGENTS.md, CLAUDE.md, rules files) and tool-server settings (such as MCP configs). A line hidden in one of them, or a risky server setting, can change what the agent does. Paste the file and the tool lists the lines and settings worth a second look, each with the line number, why it matters and a fix. The report can be copied as Markdown or downloaded as SARIF 2.1.0 (rule ids, levels and lines, with the closest CWE where one fits). You can open or drop a file, switch single rules off, skip code fences, and copy the text without its hidden characters.
The tool tries to parse the text as JSON first. If that works it is treated as a configuration, otherwise as instruction text. You can also choose the type yourself.
What it checks
Instruction text:
- hidden characters, with the count and the line and column of each: Unicode tag characters (and the hidden message they spell), direction override and isolate controls, direction marks, zero-width and invisible characters, variation selectors and control characters
- phrases that try to override earlier instructions or to hide things from the user (
ignore previous instructions,do not tell the user,hide this from the userand similar) - instructions to send data to a URL, and
curl,wgetornccommands that send file contents - long base64 or hex blocks (decoded when they hold readable text)
- commands that download and run code, such as
curl ... | sh - all URL hosts, in a table, with a closer look at raw IP addresses, cloud metadata addresses, link shorteners, paste and transfer sites, request catchers, tunnels and webhook addresses
- HTML comments and Markdown comments that address an AI model, contain a URL or a command (they are not shown when the file is displayed, but a model reads them)
- lines that tell an agent to read or include secret files (SSH keys, cloud credentials,
.env), unless the line says not to
JSON configuration (each server under mcpServers, servers or context_servers):
- a command run through a shell (
sh -c,bash -c,cmd /c, PowerShell), also when it is wrapped inenv,sudo,nohuportimeout - commands and arguments that download and run code, send files out, use
sudoor a privileged container, or delete widely (rm -rf /) - the same key written twice in one object (most programs keep the last value, a reader sees the first)
npx,uvx,pipxand similar runs with no exact version, git addresses with no commit or tag, and container images with no tag or withlatest- remote
http://orws://addresses - secrets in
env, headers and arguments, found by their format or by the name of the setting - your home folder or the root of the disk passed as an argument
- settings and flags with names such as skip-permissions, yolo, auto-approve, no-sandbox or dangerously
- wildcard allow lists such as
*orBash(*) - agent settings:
defaultMode: bypassPermissions,enableAllProjectMcpServers,trust: true, and the command of each hook
The tool does not assume one vendor's layout. It lists which keys it recognized, and says so when none were found.
What it cannot catch
- Whether something is really a problem. A security guide may quote “ignore previous instructions” on purpose, and a shell command may be fine. Every finding is a heuristic.
- Wording it does not know. A new phrasing, another language or an instruction split across several files is not matched.
- What a server or package does. The tool checks how it is started, not the code that runs. A pinned version can still be malicious.
- Other ways to run code. Container images and git addresses are checked only for a fixed tag or commit, not for what they contain. Local scripts are not read.
- YAML and TOML configs. Server settings are read from JSON only. YAML or TOML is scanned as text, and the tool says so.
- Secrets with no recognisable shape under a harmless name.
- Files that are loaded from elsewhere (includes, imports, remote prompts) are not read.
Limits
Input is limited to 100,000 characters. The check runs on the page. At most 500 findings are shown. JSON with comments or trailing commas is accepted and the tool says so. Invalid JSON is reported with the line and column of the problem, and no rules are run on it. Files you open or drop are read on the device and limited to 1 MB before the character limit applies. The file itself cannot switch rules off: an attacker could add such a marker to hide a finding, so only your own settings on this page can.
Privacy
This tool runs entirely in your browser. What you paste is processed on your device and is not sent to a server. The page does not write your input to local storage, cookies or the URL, and the tool does not send your input anywhere. The text is read as plain text only and is never run or shown as HTML.