The MCP Gateway & Registry provides a unified control plane for managing MCP servers, Artificial Intelligence agents, and reusable skills in enterprise environments. It centralizes access to tools used by autonomous agents and coding assistants, replacing ad hoc MCP server configurations with a single gateway and registry for discovery, authentication, and governance. The platform integrates with Anthropic’s MCP Registry and other external registries, supports agent-to-agent communication via an A2A protocol, and offers cross-protocol agent discovery so agents can find and call other agents based on capabilities rather than hardcoded connections.
Governance and security are core design goals, with multi-provider identity and access management that works with Keycloak and Microsoft Entra ID, plus support for Amazon Cognito and other OAuth 2.0 providers. Users and service accounts can authenticate via OAuth 2LO/3LO, session-based flows, or machine-to-machine client credentials, and the system exposes a harmonized IAM API for managing users, groups, scopes, and service accounts. Fine-grained permissions allow tool-level and method-level control, temporary access, and group-based rules that cover both MCP servers and A2A agents, while custom metadata on servers and agents enables semantic search for tags such as ownership, compliance labels, and cost centers.
The registry includes security scanning across multiple surfaces, including MCP servers, A2A agents, and skills, using Cisco AI Defense scanners with YARA pattern matching, specification validation, heuristic threat detection, and static analysis. Scan results appear directly in the UI on server and agent cards, unsafe resources can be automatically tagged or blocked, and servers with security issues are moved into a security-pending state until resolved. Compliance-oriented features include comprehensive audit logging that records all registry API and gateway access events with identity details, masked credentials, TTL-based retention with a default of 7 days, and an admin-only viewer with filtering and export to JSONL or CSV for SOC 2 and GDPR-aligned tracking of who, what, when, where, and outcome.
To support scale and distributed teams, the registry implements peer-to-peer federation between instances, as well as integration with external registries such as Anthropic MCP Registry and Workday ASOR. Federation covers MCP servers and agents with configurable sync modes, path namespacing, static token authentication, Fernet-encrypted credential storage, and generation-based orphan detection, and synced entities are read-only with source markers in the UI. Registry federation is paired with a unified semantic search API that spans servers, tools, and agents, and with a CLI and Agentic CLI that allow both humans and Artificial Intelligence agents to query the registry in natural language, discover tools, view costs and token usage, and execute MCP commands conversationally.
The platform extends the MCP model with virtual MCP servers that aggregate tools, resources, and prompts from multiple backends into a single curated endpoint. Clients connect to one virtual server while the gateway handles session multiplexing to N backend sessions, 60-second cached aggregation for tools/list, resources/list, and prompts/list, tool aliasing to resolve naming conflicts, and version pinning to lock clients to specific backend versions. Version routing features allow multiple versions of a given MCP server to run behind a single endpoint, with inactive versions tested using the X-MCP-Server-Version header, then promoted or rolled back via the API or UI, and only the active version appears in search and health checks.
Agent skills are treated as first-class resources, with a Skills Registry that discovers and governs SKILL.md files hosted on GitHub, GitLab, or Bitbucket. Skills carry YAML frontmatter metadata, support health monitoring via URL checks, visibility controls such as public, private, and group-scoped access, and star ratings and semantic search to help teams find workflows. The registry validates tool dependencies, performs automatic security scanning on skills, and protects against SSRF with redirect validation. A unified UI registration flow simplifies adding MCP servers and A2A agents, while an agent rating system with a 5-star widget, rotating rating buffer and float averages enables community-driven quality assessment of servers and agents.
On the infrastructure side, the project is container-native and targets Amazon EC2 and AWS ECS Fargate, with Terraform-based deployment that sets up multi-AZ architecture, Application Load Balancer with HTTPS and ACM, auto-scaling, CloudWatch integration, NAT Gateway high availability, Keycloak on RDS Aurora PostgreSQL, EFS shared storage, and AWS Cloud Map service discovery. File-based storage is deprecated in favor of DocumentDB and MongoDB Community Edition 8.2, which share a repository abstraction layer, provide automatic collection management and indexes, and support HNSW vector search in DocumentDB and application-level vector search in MongoDB CE, and switching between development and production backends is controlled with a single environment variable.
The project emphasizes reliability and developer workflow with a pytest-based test suite that includes 701+ passing tests across unit, integration, and end-to-end categories, a minimum coverage goal of 35% targeting 80%, and approximate 30 second execution using 8 parallel workers. Real-time metrics and observability are exposed via Grafana dashboards using SQLite and OpenTelemetry, monitoring authentication events, tool executions, discovery queries, and system performance. A management API with a Python client replaces legacy shell scripts for server, group, and user lifecycle operations, and pre-built Docker images, Podman support, and a quick start script allow teams to bring up the system quickly on developer machines or in production environments with a choice of deployment modes and registry-only configurations tailored to skills-only, MCP-server-only, agents-only, or full registry use cases.
