Hey Product Hunt 👋 I build things. Invigilate is the one I kept wishing existed. The problem: every MCP server exposes everything to the LLM by default. Every tool. Every capability. Read, write, delete — all of it sitting in the context window, consuming tokens, and one bad agent decision away from something irreversible. So I built a proxy that sits between your agent and your MCP server. One URL change. Zero modifications to your server or your agent code. Here's what it actually does: Tool Filtering — control exactly which tools the LLM can see. Hide dangerous ones entirely. Rename or redescribe tools without touching the server. Went from 50k tokens of tool definitions down to 8k in our own setup. Rules Engine — this is the core. Define conditions on any tool call — tool name, argument values, argument patterns — and set what happens when they match. Block the call outright. Send a Slack/email alert. Rules are evaluated top-to-bottom by priority, and every match is logged. Schema Sanitizer — Pydantic generates $ref/$defs in JSON schemas that break AWS Bedrock and Azure OpenAI. Invigilate auto-fixes this per provider, strips unsupported keys like title and $schema, and validates the output before it ever reaches your LLM. Always on, zero config. Audit Logs — every tool call gets a full record: tool name, arguments, matched rule, outcome (allowed / blocked / notified / pending), and timestamp. Exportable to CSV. Notifications — email and webhook (Slack-compatible) alerts fire in milliseconds when a rule matches. The payload includes the tool, args, and which rule triggered it. The thing I'm most proud of: it speaks standard JSON-RPC 2.0, so it's completely invisible to both sides. Neither your agent nor your MCP server knows it's there. Would love to know what MCP servers you're running and what you wish you could control about them. Ask me anything 👇

Invigilate

Your agents, under control

Your agents, under control