Let AI agents use production data without handing them your database

Short summary

Synapsor Runner is an Apache-2.0 runtime that sits between MCP clients and PostgreSQL/MySQL, exposing reviewed semantic capabilities instead of raw SQL to AI agents. The model can read only contract-allowed columns and rows, propose changes through a proposal-evidence-replay boundary, but never receives commit or write authority—all approvals and writebacks happen outside the model loop. Testing with injection-attack scenarios on real LLM agents showed zero cross-tenant reads and zero unauthorized writes, because enforcement lives server-side rather than relying on model compliance.

• •Exposes semantic capabilities (e.g., billing.inspect_invoice) instead of raw SQL to AI agents, eliminating model-facing write tools
• •Enforces tenant scope, allowed columns, and row limits via server-side contracts; proposals require external approval before any writeback
• •Testing with prompt-injection attacks yielded zero cross-tenant reads and zero unauthorized writes; also reduces token usage by avoiding schema dumps and SQL retry loops

Generated with AI, which can make mistakes.