// the find
hoophq/hoop
One gateway in front of every protocol. Same policy across MCP, LLMs, databases and containers. Wire-level enforcement at under 5ms.
hoop is a layer 7 gateway that sits between engineers or AI agents and production infrastructure — databases, Kubernetes, SSH, RDP — parsing wire protocols in real time to mask PII before it leaves, block destructive commands before they execute, and record every session for replay. It's aimed at teams where developers or autonomous agents touch production data and compliance evidence is required. Written in Go with a Rust sidecar for the PII detection and OCR pipeline.
Wire-protocol parsing at this breadth (PostgreSQL, MySQL, MSSQL, MongoDB, SSH, K8s, HTTP/gRPC, RDP) is genuinely hard work, and they have integration tests with real containers for each protocol — not just happy-path smoke tests. The Rust PII gate (`agentrs/src/piigate/`) uses ML classification rather than regex, which means it handles contextual cases like distinguishing a phone number in a `phone` column from a build ID with the same digit pattern. The MCP gateway angle is well-timed: as AI agents start running database queries autonomously, having a proxy that masks JSON payloads before they reach the model and blocks destructive tool calls is a real gap. The session recording with full replay is more useful than audit logs — you can reconstruct exactly what an agent saw, not just what it asked for.
769 stars against claims of NYSE-listed customers and 5,000 databases protected is a suspicious gap — either the open-source build is a stripped-down lead-gen vehicle for a paid tier not described in the README, or the traction numbers are marketing. The operational complexity is real but undersold: you're deploying Go gateway + Rust binary + OCR sidecar + identity provider integration, and the Dockerfiles alone (five of them, including GPU variants) hint at what production actually looks like. The 'Agent in the loop' risk scoring was shipped June 2026 and uses an LLM to evaluate the risk of AI agent actions in real time — this adds latency and cost that gets zero quantification beyond 'under 5ms' for the gateway itself (the LLM call is conveniently excluded). With 56 forks and no visible community contributions beyond the core team, this is effectively a product with public source, not an open-source community project.