Solving Agent Write-Access Risks: Open-Source Security Gateway 'mcpip'
Ok_Anxiety410888 · reddit · 2026-08-05
The author argues that read-only agents are easy, but granting write access (sending emails, updating DBs) exposes flaws in simple approval prompts, primarily 'approval fatigue.' To fix this, they built mcpip, a security gateway sitting in front of tool calls.
Core Design Principles:
- Issued Tool Lists: The agent only sees tools it's identity-allowed to see, preventing prompt injection from reaching for unauthorized tools.
- Hidden Real Targets: The agent calls abstract actions (e.g., skillpayrollrun) without ever seeing real connection strings or credentials, preventing leaks.
- Payload-bound Approval: Approval tokens are tied to exact request bytes. You can't approve a $50 transfer and use the token for a $5,000 one; changing a field invalidates the approval.
- Pre-execution Logging: Ensures audit logs exist even if the agent crashes mid-execution.
The gateway runs deterministically (no LLM inference), speaks MCP, and is self-hostable.
Related event: Open-Source Gateway mcip Secures AI Agent Tool Calls(2 posts)→
More from coding & agent
- AI Assists $15 Hardware Build, But Fails at PCB Layout — johnowhitaker · 2026-08-05
- API Key Expiry Leads to Runaway Agent, Costs $300 in Idle Compute — voooooogel · 2026-08-05
- Dev Builds Pixel-Art GPU Cluster Dashboard in 20 Mins Using GLM Agent — Porespellar · 2026-08-05
- Open-Source AI Children's Picture Book Workflow with Local Voice Cloning — sujingshen · 2026-08-05
- Building In-Browser Coding Agents: Verification is the Main Bottleneck — Death12th · 2026-08-05
- Agent Skills Ecosystem Booms: skills.sh Launches Open Directory and Leaderboard — shadcn · 2026-08-05