DARINORCold Lab

Tools / Agent Config Checker

AI Agent Config Checker

Paste a system prompt or agent config (Claude Code settings, MCP config, OpenAI assistant JSON, or plain markdown) and check for leaked secrets, over-broad tool permissions, and risky directives. 100% in-browser.

Everything runs locally — your config never leaves your browser. Static heuristics only; this is a first-pass triage, not a scanner.

Findings
Waiting for input

Paste a config to scan. Try the sample to see what the checker catches.

// About this tool

AI Agent Config Checker

AI agents run with elevated permissions, and the configs that grant them are often committed to repos or pasted into prompts with secrets inside. This checker scans a pasted config, system prompt, or settings JSON for the patterns that actually get agents compromised.

It looks for three classes of risk: leaked credentials (API keys, tokens, private keys, database URIs), over-broad permissions (auto-approval, wildcard tool grants, permission bypass), and instruction-like directives in content-reading positions — the raw material of tool-poisoning and prompt-injection attacks.

The scan runs entirely in your browser. Nothing is uploaded, which matters because the thing you're scanning is usually itself sensitive. Findings include a suggested fix for each issue.

This is a first-pass triage tool, not a scanner. Static checks cannot detect semantic prompt injection — if an attacker's intent is disguised well enough, no heuristic will catch it. Use the results as a starting point, not a clean bill of health.

// When to use it

  • Audit a .claude/settings.json before committing

    Paste the file and catch a leaked API key or an alwaysAllow block before it lands in a repository.

  • Check an .mcp.json for risky tool grants

    Spot wildcard tool lists and auto-approve settings that would let a prompt-injected agent act without a human gate.

  • Review a system prompt for injection-friendly directives

    Flag instruction-like phrasing in descriptions of content-reading tools — a common tool-poisoning pattern.

// Questions

Is my config uploaded anywhere?

No. The scan runs entirely in your browser using static heuristics. Nothing is sent to a server, which is deliberate — the config you're scanning is usually itself sensitive.

What does the checker detect?

Credential patterns (API keys, tokens, private keys, database URIs with passwords), permission issues (auto-approval, wildcard grants, permission bypass), and instruction-like directives in content-reading positions.

Can it detect prompt injection?

Not reliably, and no static tool can. Prompt injection is a semantic attack — instructions and data share one channel. The checker flags obvious directive patterns in risky positions, but treat a clean result as a starting point, not a clean bill of health.

Which config formats does it accept?

JSON configs (Claude Code settings, MCP configs, OpenAI assistant JSON), YAML-style key-value text, and plain markdown prompts. The JSON path gets deeper structural checks; text formats get pattern matching.

// Related tools

This site uses minimal cookies and local storage to keep features like the chat widget and games working. We do not use third-party tracking cookies. Privacy Policy