Glyph

Read the runes before your agent steps on them.

[](https://opensource.org/licenses/MIT) [](https://www.python.org/downloads/) []()

An MCP security scanner that finds tool poisoning, credential leaks, and insecure transports in your AI agent configurations — before attackers do.

---

MCP (Model Context Protocol) is how AI agents connect to tools. Claude Desktop, Cursor, pi, Windsurf — they all use it. There are now 16,000+ MCP servers in the ecosystem.

66% of them have security findings.

Tool poisoning — hiding malicious instructions in tool descriptions — has a 91% success rate against production AI agents. The first malicious MCP server was found on npm in September 2025, silently BCC'ing every email to an attacker.

Nobody audits these servers before connecting. Glyph does.

---

🔴 Tool Poisoning Detection — Finds hidden instructions, prompt injection, unicode tricks, base64 payloads, and behavioral directives buried in tool descriptions

🔴 Credential Exposure — Catches hardcoded API keys (OpenAI, AWS, GitHub, Anthropic), tokens, and secrets that should be in environment variables

🟡 Transport Security — Flags MCP servers running over plain HTTP, missing TLS, or lacking authentication

🔵 Live Scanning — Connects to real MCP servers via JSON-RPC, pulls actual tool definitions, and scans them in real time

---

pip install glyph-scan

Scan a config file:

glyph scan ~/.config/claude/claude_desktop_config.json

That's it. Results in seconds.

---