Your agent connects to MCP servers — and everything it does stays inside the permissions you granted.
So when something goes wrong — a bad instruction, a compromised tool, a prompt injection — the damage is already bounded.
Works with Claude Code, Codex, and any MCP agent, via a PreToolUse hook.
The browser core disagreed with the Node core this page claims to be. Rather than render a demo that would be lying about what it runs, here is the divergence:
Watch an unauthorized tool call get stopped
The agent holds your credentials directly. Nothing evaluates what it asks for.
what TeleKey handed the sub-agent
what the agent actually got to do
Same request, same injection. Every call now goes through the engine first.
what TeleKey handed the sub-agent
checks every call before it runs
what the agent actually got to do
Every line below came from a real call
| Requested capability | Verdict | Why |
|---|
This repo is a plugin marketplace. Point Claude Code at it — no clone, no build.
/plugin marketplace add stefans71/telekey-mcp
Scope & honesty
Enforcement bounds what a hijacked agent can do; it does not stop the hijack. Signing here is HMAC-SHA256 via Web Crypto — a stand-in for RFC 9068 JWTs, exactly as in the Node reference.