Securing Agent Tool Calls: Open-Source MCP Gateway mcip
Ok_Anxiety410888 · reddit · 2026-08-05
A developer open-sourced mcip, a self-hosted security gate for AI agent tool calls, designed to mitigate risks where agent skills might bypass the LLM's own safety checks.
Core Mechanics
- Per-call Interception: Abandons easily ignored approval prompts in favor of policy checks on every single tool call.
- Opaque Denials: Denials only return a correlation ID, preventing a probing agent from learning system details from error messages.
- Payload-bound Approval: High-risk actions (e.g., paying $50) require a code bound to that exact payload, preventing malicious amount modifications (e.g., executing $5,000).
Technical Features
- Speaks MCP natively, acting as a gatekeeper for existing MCP clients without extra code.
- No model inference in the decision path; never sees user prompts or API keys.
- Includes offline tests with 22 attack paths and 7 allow-paths, encouraging white-hat breaking.
Related event: Open-Source Gateway mcip Secures AI Agent Tool Calls(2 posts)→
More from coding & agent
- Maximizing 8GB VRAM: A Hybrid Local and Frontier Model Coding Workflow — RootExploit_ · 2026-08-05
- Developer Shares 4 Practical AI Agent Skills for Better Workflows — smtabatabaie · 2026-08-05
- Stanford's Self-Improving AI Agents Course Now Available Online — salgar · 2026-08-05
- Voice Assistant Bottlenecks: ASR Diarization and TTS Latency — dangerous_inference · 2026-08-05
- AI Agent Experiment Shut Down Due to High Maintenance Overhead — every · 2026-08-05
- Company Shuts Down Internal AI Agents: Maintenance Costs Negate Efficiency Gains — every · 2026-08-05