An enterprise-grade AI security gateway that unifies proxying, auditing, and governing all AI API calls and MCP tool invocations across an organization.
ThinkWatch is an enterprise-grade AI security gateway built with a Rust backend and React frontend. It employs a dual-port architecture (public Gateway :3000 + internal Console :3001) to uniformly govern all AI API and MCP tool invocations within an organization.
AI API Gateway: Natively supports OpenAI Chat Completions, Anthropic Messages, and OpenAI Responses formats on a single port. Multi-provider routing across OpenAI, Anthropic, Gemini, Azure OpenAI, AWS Bedrock, and any OpenAI-compatible endpoint with automatic format conversion. Model-prefix-based auto-routing, zero-overhead SSE streaming with real-time token counting, virtual API Key (tw- prefix) lifecycle management, multi-dimensional sliding-window rate limits and natural-period token budgets, model-level token weighting, three-state circuit breaker with exponential backoff retry, and real-time cost tracking.
MCP Gateway: Aggregates all upstream MCP server tools into a single endpoint with namespace isolation (e.g., github__create_issue), tool-level RBAC, connection pooling with health monitoring, and full audit logging for every tool invocation.
Security & Compliance: 5-level RBAC (Super Admin → Viewer), SSO/OIDC integration, AES-256-GCM encryption at rest, SHA-256 API Key hashing, CSP headers, JWT entropy enforcement, session IP binding, automatic PII redaction, and a 2MB Distroless container image.
Observability: Prometheus metrics endpoint, enhanced health checks, ClickHouse columnar audit logs, Syslog/Kafka/Webhook log forwarding, and a unified log browser.
Supports Docker Compose and Kubernetes Helm Chart deployment. Works as a plug-and-play replacement backend for Cursor, Claude Code, Cline, and other AI-powered tools. Current version v0.1.0 with 555+ commits, licensed under BSL 1.1 (Change Date 2030-04-02, then GPL-2.0-or-later).