Semantic GPS
Semantic GPS is the governance gateway between AI agents and the business systems they were never supposed to touch unsupervised.
Core trick
Proxies every MCP tool call through a policy gateway that runs in shadow mode first, then enforces and logs each call.
Description
Semantic GPS is the governance gateway between AI agents and the business systems they were never supposed to touch unsupervised.
It sits between any MCP-connected agent and any vendor MCP server as one control plane: shadow → enforce live policy mode swap, audit trail on every call with grouped trace IDs, saga rollback with explicit per-step input mapping, and a Tool Relationship (TRel) MCP extension so agents can discover safe workflow paths instead of guessing.
Twelve gateway-native policies span seven governance dimensions including time gates, identity, residency, data hygiene, and idempotency.
A side-by-side Playground proves the contrast between raw and governed under identical Opus 4.7 (Or bring your own LLM via APIkey) prompts.
Three of the EU AI Act's hardest operational obligations:
- Article 9 (risk management),
- Article 12 (record-keeping)
- Article 14 (human oversight)
map directly onto primitives Semantic GPS tries to ship, ahead of the August 2, 2026 enforcement date. The recent Replit, Meta, and Cursor incidents all happened at the same seam: agent reaches business system unsupervised. Semantic GPS could become the missing layer.
README excerpt
# Semantic GPS **The governance gateway between AI agents and the business systems they were never supposed to touch unsupervised.**  **Recent receipts.** February 2026: Meta's Director of AI Alignment lost 200+ emails to an autonomous agent that ignored her stop commands. July 2025: Replit's AI deleted a live production database during a declared code freeze and lied about recovery. 97% of enterprises expect a major AI-agent incident within the next 12 months; only 14.4% of agents reach production with full security review. The agents are working. The safety surface around them isn't. Semantic GPS sits between agents and any MCP-connected system as one control plane: shadow→enforce live policy swap (the observe-before-acting pattern those incidents lacked), audit on every call, saga rollback with explicit per-step input mapping, a Tool Relationship (TRel) MCP extension for workflow discovery, and a side-by-side Playground proving raw-MCP vs governed contrast under identical Opus 4.7 prompts. Built for the Anthropic "Keep Thinking" Hackathon (April 2026). 5-day scope. - **Live demo:** https://semantic-gps-hackathon.vercel.app/ - **Demo video:** https://www.youtube.com/watch?v=fYh2MpMw1ng - **Full story:** [`docs/VISION.md`](./docs/VISION.md) - **Security:** [`SECURITY.md`](./SECURITY.md) --- ## What shipped - **Vendor-agnostic MCP gateway.** Customers register their own MCP servers via `POST /api/servers` (HTTP-