Project Name
MCP Gateway & Registry
Project Description
MCP Gateway & Registry is a unified platform for centralizing access to AI Assets, which today include three categories: MCP (Model Context Protocol) servers, AI agents, and agent skills. The platform is designed so additional AI asset types can be added over time without restructuring the governance model. It serves four core functions: (1) a unified MCP Server Gateway providing a single access point for multiple MCP servers, (2) an MCP Servers Registry for registering, discovering, and governing MCP server access, (3) an Agent Registry and A2A Communication Hub for agent registration, discovery, governance, and direct agent-to-agent communication through the A2A protocol, and (4) an Agent Skills Registry for registering, discovering, and governing reusable agent skills, with per-skill authentication credentials and integrated security scanning.
The project originated in May 2025 to solve a growing problem in enterprise AI development: as teams adopt AI coding assistants (VS Code, Cursor, Claude Code) and deploy autonomous AI agents, each developer and agent must independently configure connections to dozens of MCP servers, manage scattered credentials, and operate without centralized governance. This creates security risks, credential sprawl, zero visibility into tool usage, and no standard for agent-to-agent discovery.
The platform replaces this chaos with a single control plane. Developers, AI coding assistants (VS Code, Cursor, Claude Code, and others), and AI agents connect to one gateway and gain access to all approved MCP servers with enterprise authentication (Keycloak, Microsoft Entra ID, Okta, Auth0, Amazon Cognito), fine-grained access control at the tool and method level, comprehensive audit logging, and real-time observability. Key capabilities include dynamic tool discovery via semantic search, virtual MCP servers (tool aggregation from multiple backends), peer-to-peer registry federation, agent skills registry, MCP server security scanning (via Cisco AI Defense), A2A protocol support, and integration with external registries (Anthropic MCP Registry, AWS Agent Registry / Amazon Bedrock AgentCore).
As of release 1.23.0, the project has 642+ GitHub stars, 165+ forks, 37+ contributors, 700+ passing tests, and has been deployed on AWS ECS, EKS, Docker Compose, and Podman across multiple organizations. An AWS Workshop Studio lab ("Securing AI Agent Ecosystems with MCP Gateway and Registry") is available for hands-on learning.
Alignment with AAIF Mission
MCP Gateway & Registry directly advances the Agentic AI Foundation's mission of fostering open, interoperable, and well-governed agentic AI systems:
-
Open Standards Adoption: The project is a production gateway and registry layer built around two foundational agentic AI standards: the Model Context Protocol (MCP) and the Agent-to-Agent (A2A) protocol. For MCP, the project acts as a gateway (proxying traffic to registered MCP servers) and implements the Anthropic MCP Registry REST API v0.1 specification so any MCP-compliant client can consume servers from the registry without modification. For A2A, the project is a registry-only layer that hosts A2A agent cards and provides authenticated discovery, while agents continue to communicate peer-to-peer using the A2A protocol (the project does not proxy or reimplement the A2A wire protocol). The project does not define new standards; it adopts these existing ones and adds the governance and discovery layer around them.
-
Interoperability and Federation: The platform was designed to avoid registry silos. It supports three federation modes: (a) peer-to-peer federation between independent MCP Gateway & Registry instances (so an agent registered in one enterprise's deployment can be discovered by a partner running their own deployment), (b) imports of upstream servers from the Anthropic MCP Registry, and (c) federation with AWS Agent Registry / Amazon Bedrock AgentCore. This ensures agents and tools remain portable across vendors and across organizational boundaries.
-
Governance and Trust: The project provides the governance layer enterprise adoption of agentic AI requires: fine-grained access control at the tool and method level, compliance audit logging, integrated security scanning of MCP servers, A2A agents, and agent skills (via Cisco AI Defense scanners), centralized credential management, and support for five enterprise identity providers (Keycloak, Microsoft Entra ID, Okta, Auth0, Amazon Cognito). This directly supports responsible AI agent deployment in regulated industries.
-
Community-Driven Development: The project is Apache 2.0 licensed with 37+ contributors from multiple organizations, 21+ releases in 12 months, comprehensive documentation, and an AWS Workshop Studio lab used by participants worldwide to learn secure agentic AI deployment patterns.
Relation to Existing AAIF Projects
MCP Gateway & Registry is designed to complement and extend the three projects currently hosted by AAIF (as listed in the AAIF landscape):
-
Model Context Protocol (MCP): This is the most direct relationship. MCP Gateway & Registry is a production governance, discovery, and gateway layer for the MCP ecosystem. It implements the Anthropic MCP Registry REST API specification so that any MCP-compliant client can consume servers from the registry without modification, imports servers from the upstream Anthropic MCP Registry, and applies enterprise policy (authentication, fine-grained authorization, audit logging, security scanning via Cisco AI Defense) before MCP servers are exposed to developers or AI agents. Where MCP defines the protocol, this project supplies the scaled, governed deployment layer that enterprise MCP adopters need.
-
goose: goose is a client-side extensible AI agent that connects to MCP servers. MCP Gateway & Registry is directly useful to goose users and operators: instead of each goose instance maintaining its own list of MCP server connections and credentials, goose can point at a single MCP Gateway endpoint and inherit the registry's governed catalog, authentication, and tool-level authorization. An initial goose integration already exists: the registry UI exposes goose as a first-class connection option alongside VS Code, Cursor, and Claude Code (see merged PR #1018), so goose users can connect to a governed MCP Gateway with a single click. We see goose + MCP Gateway & Registry as a natural "agent runtime + agent infrastructure" pairing for AAIF users, with room to deepen the integration over time.
-
AGENTS.md: AGENTS.md provides a standardized, predictable file format for giving context and instructions to AI coding agents inside a project. This project and AGENTS.md are complementary at different layers: AGENTS.md scopes per-repository agent instructions; MCP Gateway & Registry scopes organization-wide tool discovery, identity, and governance. A coding agent following AGENTS.md conventions can discover and invoke the right MCP tools (and A2A agents) through an AAIF-governed registry rather than hardcoded configuration.
Open standards the project implements (not AAIF-hosted, included here for context):
-
Agent-to-Agent (A2A) Protocol: The project provides a production-grade registry and discovery layer for A2A-capable agents. A2A defines the communication protocol; MCP Gateway & Registry solves the discovery and governance problem: how agents find each other, authenticate, and operate within organizational policies. Agents register A2A agent cards in the registry and discover other agents via semantic search, then communicate directly peer-to-peer using A2A. Supporting A2A alongside MCP is intentional: it keeps the governance layer neutral across the two dominant agentic interoperability standards.
-
Anthropic MCP Registry REST API: The project implements this upstream specification so that the MCP ecosystem's existing client and tooling base works against this registry unchanged.
Complementary candidate projects currently under AAIF review. Several projects with active proposals on aaif/project-proposals are natural architectural complements rather than competitors. Envoy AI Gateway (proposal #18) operates at the network/proxy layer for LLM and MCP traffic (token-based rate limiting, multi-provider routing, upstream authentication); MCP Gateway & Registry operates at the discovery, catalog, and identity-aware governance layer. A production deployment could run both: Envoy AI Gateway as the data-plane proxy, MCP Gateway & Registry as the control-plane catalog. Similarly, agentgateway (proposal #11) and policy-focused proposals like DAP (#17) and PIC Standard (#16) address overlapping governance concerns at adjacent layers. We would welcome joint architectural discussions under an AAIF umbrella if any of these are accepted.
Example Use Cases and Evidence of Adoption
Use Case 1: Enterprise AI Coding Assistant Governance
Organizations with hundreds of developers using AI coding assistants (VS Code with Copilot, Cursor, Claude Code) deploy MCP Gateway & Registry to provide a single, governed configuration point. Instead of each developer managing their own MCP server connections and credentials, central IT manages a curated catalog of approved MCP servers accessible through the gateway with SSO authentication and fine-grained permissions.
Use Case 2: Autonomous Agent Ecosystems
Teams building multi-agent systems use the registry as a discovery and governance layer. Agents register their capabilities and discover other agents through semantic search, then communicate directly via the A2A protocol. The gateway provides M2M authentication (OAuth2 Client Credentials), access control, and audit trails for all agent interactions.
Use Case 3: AWS Workshop Studio Deployment
The project is featured in an official AWS Workshop Studio lab ("Securing AI Agent Ecosystems with MCP Gateway and Registry") where participants deploy the full stack on AWS, configure enterprise authentication, implement access control, and secure AI agent communications. This demonstrates production-readiness and educational value.
Evidence of Production Deployments:
- A global online travel platform: Production deployment for AI agent and MCP server governance
- A global retirement and insurance services company: Production deployment for enterprise AI tooling infrastructure
Telemetry-Backed Adoption Data (as of 2026-05-11):
The project collects anonymous, opt-out telemetry (no PII, no hostnames). The following metrics are from the collection window ending 2026-05-11:
- 349 unique registry instances deployed over the collection period (lifetime count).
- 15 deployments confirmed actively running for 14+ days (defined as 10 or more daily heartbeats within the trailing 14-day window), providing a conservative floor on sustained production-style usage. An additional 107 non-internal instances are sustained for 3+ days.
- New-install rate averages approximately 20 per day over the last 7 days (range: 5 to 29 new unique registry instances per day), indicating continuous, ongoing adoption rather than a one-time spike.
- Multi-cloud adoption: AWS dominant, with both Azure and GCP represented among confirmed-alive instances.
- Enterprise authentication in use: Okta, Microsoft Entra ID, Keycloak, Amazon Cognito, and Auth0 all observed in production deployments, confirming enterprise-grade usage.
- Production compute platforms: Kubernetes, AWS ECS, and Docker all represented, with no bare-VM deployments (every deployment is containerized or orchestrated).
- Real-world scale: The most feature-rich customer deployments register 66 MCP servers and 34 agents, running on Kubernetes with Okta authentication.
- Active usage: 117,057 total lifetime semantic search queries across the fleet, demonstrating that deployed registries are actively used for tool and agent discovery.
Community Signals:
- 642+ GitHub stars and 165+ forks
- 37+ contributors from multiple organizations
- Featured AWS Workshop Studio lab with production deployment patterns
- Featured in the AWS Show & Tell presentation on YouTube
- Amazon ECR Public pre-built multi-architecture images (
public.ecr.aws/p3v1o3c6/) with regular releases
- Helm charts for Kubernetes/EKS deployment
- Active GitHub Discussions and Issues with community engagement
Technical Committee Sponsor (if identified)
James Ward, Principal Developer Advocate, AWS.
GitHub Repository URL
https://github.com/agentic-community/mcp-gateway-registry
License
Apache 2.0
Governance Model
The project currently follows a maintainer-led governance model:
- Core Maintainers: Amit Arora (project lead) and a small group of core maintainers with commit access who review and merge pull requests.
- Decision Making: Technical decisions are made through GitHub Issues and Pull Requests with open discussion. Major architectural decisions are documented in the
docs/design/ directory with design documents.
- Contribution Process: Documented in CONTRIBUTING.md, following a fork-and-PR workflow with code review requirements.
- Code of Conduct: The project has adopted the Amazon Open Source Code of Conduct.
- Security Policy: Responsible disclosure process documented in SECURITY.md.
- Community Communication: Today, community discussion happens primarily on GitHub Issues and GitHub Discussions, with design documents in
docs/design/ and regular release notes in release-notes/. The project commits to launching a public bi-weekly community meeting with open attendance and published meeting notes within 30 days of AAIF acceptance, to provide a synchronous venue for roadmap discussion, contributor onboarding, and cross-project coordination with other AAIF projects.
The project is prepared to adopt LF Minimum Viable Governance or the AAIF governance framework upon acceptance as a hosted project.
CI/CD & Release Workflow
The project uses GitHub Actions for CI/CD with 12 workflow files covering:
- Testing:
registry-test.yml runs the full pytest suite (701+ tests) with 8 parallel workers on every PR targeting main/develop. Separate workflows test the auth server (auth-server-test.yml), Helm charts (helm-test.yml), Terraform configs (terraform-test.yml), and metrics service (metrics-service-test.yml).
- Building: Dedicated build workflows for each component:
build-registry.yml, build-auth-server.yml, build-mcpgw.yml build and push multi-architecture container images to Amazon ECR Public (public.ecr.aws/p3v1o3c6/).
- Release Process:
release-images.yml builds multi-architecture (AMD64/ARM64) Docker images and pushes tagged releases. Helm chart updates are automated via helm-chart-update.yml and helm-release-retag.yml.
- Documentation:
docs.yml builds and deploys documentation to GitHub Pages.
- Pre-commit Hooks: Automated code quality checks including ruff (linting/formatting), bandit (security scanning), mypy (type checking), and pytest (fast unit tests).
Release Frequency: The project has shipped 21 releases since May 2025 (roughly biweekly to monthly), with detailed release notes in the release-notes/ directory. Each release follows semantic versioning.
Public-Facing Contribution Process for Specifications
The project's contribution process for specifications and design changes:
-
Design Documents: Major features are proposed and documented in docs/design/ with comprehensive design documents covering architecture, data models, API contracts, and security considerations. Examples include federation architecture, virtual MCP server design, hybrid search architecture, authentication design, and AWS Agent Registry federation.
-
GitHub Issues: All feature proposals start as GitHub Issues with detailed descriptions, acceptance criteria, and community discussion. The project uses milestones to group issues into releases.
-
Pull Request Review: All pull requests are reviewed and tested by one of the core maintainers before merge. Review guidelines (including test-pass requirements and a "Merge Specialist" persona for comprehensive PR evaluation) are documented in the project's CLAUDE.md.
-
Contributing Guide: CONTRIBUTING.md documents the full contribution workflow including issue filing, PR submission, and code review expectations.
Publicly Accessible Issue Tracker
https://github.com/agentic-community/mcp-gateway-registry/issues
Currently 100+ open issues with active community engagement. Across the project's ~12 month history, the repository has accumulated 400+ issues (101 open, 299 closed) and 590+ pull requests (20 open, 500+ merged), for a combined total of nearly 1,000 tracked changes. Issues are organized into per-release and per-week milestones (recent releases v1.0.20, v1.0.21, v1.0.22, and 1.23.0 are all complete; 1.24.0 and 1.25.0 are in progress) with labels for categorization. The project also uses GitHub Discussions for feature requests and general discussion.
External Project Dependencies
Key runtime dependencies (all open-source):
| Dependency |
License |
Purpose |
| FastAPI (>=0.115.12) |
MIT |
Web framework for REST API |
| Pydantic (>=2.11.3) |
MIT |
Data validation and settings |
| Uvicorn (>=0.34.2) |
BSD-3 |
ASGI server |
| Motor (>=3.3.0) / PyMongo (>=4.6.0) |
Apache 2.0 |
MongoDB async/sync drivers |
| MCP SDK (>=1.9.3) |
MIT |
Model Context Protocol client |
| FAISS-cpu (>=1.7.4) |
MIT |
Vector similarity search |
| sentence-transformers (>=3.0.0) |
Apache 2.0 |
Local embeddings |
| LiteLLM (==1.83.0) |
MIT |
Multi-provider LLM/embeddings |
| Boto3 (>=1.42.87) |
Apache 2.0 |
AWS SDK |
| LangChain Core (>=1.2.28) |
MIT |
LLM orchestration |
| LangGraph (>=0.4.3) |
MIT |
Agent orchestration |
| Strands Agents (>=0.1.6) |
Apache 2.0 |
Agent framework |
| Cisco AI MCP Scanner (>=3.0.1) |
Apache 2.0 |
MCP server security scanning |
| Cisco AI A2A Scanner |
Apache 2.0 |
A2A agent security scanning |
| Cisco AI Skill Scanner (>=1.0.0) |
Apache 2.0 |
Agent skills security scanning |
| Cryptography (>=46.0.7) |
Apache 2.0 / BSD |
Fernet encryption for federation |
| PyJWT (>=2.12.0) |
MIT |
JWT token handling |
| Prometheus Client (>=0.20.0) |
Apache 2.0 |
Metrics export |
| Rich (>=13.0.0) |
MIT |
CLI formatting |
| HTTPX (>=0.27.0) |
BSD-3 |
Async HTTP client |
Infrastructure dependencies:
- MongoDB (SSPL) or Amazon DocumentDB (proprietary): Storage backend
- Keycloak (Apache 2.0) / Okta / Microsoft Entra ID / Auth0 / Amazon Cognito: Identity providers
- NGINX (BSD-2): Reverse proxy gateway
- Docker (Apache 2.0): Container runtime
Maintainers & Contributors
Core Maintainers:
| Name |
GitHub Handle |
Affiliation |
| Amit Arora |
@aarora79 |
Amazon Web Services |
| Omri Shiv |
@omrishiv |
Amazon Web Services |
Note: The project has two AWS-employed core maintainers today. The broader contributor base spans 37+ individuals across multiple organizations. Expanding to multi-organization core maintainership is a committed post-acceptance goal with a 6-month target, promoting active contributors based on sustained technical contributions and code review participation. AAIF hosting would provide the formal governance framework (documented maintainer promotion criteria, contribution-based advancement) to make this transition credible and repeatable.
Active Contributors (37+ total, top non-maintainer contributors by commit count as of 2026-05-11):
Bot contributors (Dependabot, GitHub Actions) are excluded from the table above. Full contributor list: https://github.com/agentic-community/mcp-gateway-registry/graphs/contributors
Leadership Team & Decision Process
Project Lead: Amit Arora (@aarora79) provides overall technical direction and coordinates the roadmap.
Decision-Making Process:
- Day-to-day decisions: Made by core maintainers through PR reviews. Any core maintainer can approve and merge changes that align with project conventions.
- Feature proposals: Submitted as GitHub Issues, discussed openly, and prioritized into milestones. Major features require a design document in
docs/design/ before implementation.
- Architectural decisions: Require consensus among core maintainers. Design documents serve as the record of decision for significant changes.
- Release decisions: The project lead coordinates release timing based on milestone completion, with input from maintainers on readiness.
- Conflict resolution: Escalated to the project lead if consensus cannot be reached among maintainers.
The project is open to adopting a more formal governance structure (such as a Technical Steering Committee) as part of joining AAIF.
Roadmap
The project maintains a public roadmap organized into release milestones on GitHub:
Completed (recent):
- v1.0.20: Unified JWT+static token auth, registration gate/webhooks, multi-key API tokens, M2M registration, Python 3.14 upgrade (addressing CVE-2025-13836)
- v1.0.21: Admin data export, per-skill auth credentials, centralized log rotation, ARM64 images
- v1.0.22: Group-restricted agents, OAuth2 registration gate (client credentials), MongoDB URI override, local-only group fixes
- 1.23.0 (current, May 2026): Cloud detection (ECS meta / k8s / IMDS probes reducing
cloud=unknown telemetry bucket), IAM hardening (M2M orphan delete + deduplicated user listing), Splunk-ready JSON-Lines logging, and a switch to strict semantic versioning (no v prefix) for all tags, images, and Helm chart versions.
Planned (next 3 months):
- 1.24.0 (in progress, 9 open issues): Prometheus
/metrics endpoint for the registry, A2A reverse-proxy gateway support, dependency management for agent/service/skill cards, Registry Copilot (embedded AI chat assistant for registry operations), Helm ALB ingress-scheme configurability, UI title configurability, and MCP registration deduplication.
- 1.25.0 (scoped, 8 open issues): Coding-Assistant OAuth Integration umbrella (#988) covering PRM + Authorization Server metadata,
WWW-Authenticate flow, Entra v1 scope verbatim support, RFC 8707 enforcement, and a documented no-DCR decision.
Backlog (6-12 months):
- Agent-to-agent knowledge sharing and Context Hub MVP
- MCP OAuth 2.1 Authorization Spec (RFC 9728)
- AI Gateway and Registry rebrand (reflecting expanded scope beyond MCP)
- Multi-level tool usage rate limiting
Full roadmap with links to all issues:
Security
The project does not currently hold an OpenSSF Best Practices badge. However, the project demonstrates strong security practices:
- Automated Security Scanning: Bandit is integrated into CI/CD and pre-commit hooks for Python security vulnerability detection
- Comprehensive Security Documentation: Security Posture Guide covering defense-in-depth across ECS, EKS, and Docker Compose deployments
- Responsible Disclosure: SECURITY.md with documented disclosure process
- Integrated Security Scanning for Registered Assets: MCP servers, A2A agents, and agent skills are automatically security-scanned during registration using Cisco AI Defense scanners
- Enterprise-Grade Authentication: OAuth 2.0 with support for 5+ identity providers, fine-grained RBAC, compliance audit logging
- CVE Responsiveness: Recent upgrade to Python 3.14 to address CVE-2025-13836
Pursuing an OpenSSF Best Practices badge is a near-term goal that AAIF hosting would help accelerate.
Website URL
https://agentic-community.github.io/mcp-gateway-registry/
Documented Governance Practices (if any)
No formal GOVERNANCE.md, MAINTAINERS.md, or CODEOWNERS file exists today. The project commits to creating all three aligned with AAIF's governance framework within 6 months of acceptance. DCO (Developer Certificate of Origin) sign-off will be adopted on all commits as part of this transition, replacing the project's current Amazon Open Source Code of Conduct attestation flow with the standard AAIF/LF contribution process.
Links to Social Media Accounts
The project does not yet maintain dedicated social accounts. Project updates, new release announcements, and community posts are currently shared every 2 to 3 weeks via the project lead's personal LinkedIn account: Amit Arora. An example of a recent release post is here. On acceptance into AAIF, this will be formalized with a dedicated project LinkedIn page (and an X account) within 30 days, so that community engagement and release communications live under the project's own identity rather than an individual contributor's.
Trademark and accounts
Details of Existing Financial Sponsorship
The project receives engineering contributions from AWS, which employs the two current core maintainers (Amit Arora, Omri Shiv) and funds their contributions to the project. Pre-built multi-architecture (AMD64/ARM64) container images are published to Amazon ECR Public under public.ecr.aws/p3v1o3c6/ and referenced by the project's Helm charts. GitHub Actions provides CI/CD via GitHub-hosted runners. Documentation is hosted on GitHub Pages.
Infrastructure Needs or Requests
The project's current infrastructure is largely self-hosted or funded by AWS, but two AAIF-provided resources would materially strengthen the project:
-
Security audit tooling and OpenSSF Best Practices badging support: The project is committed to achieving an OpenSSF Best Practices badge (see Security section). Foundation-facilitated access to third-party security audit tooling, SBOM scanning infrastructure, and AAIF-community review of our security posture documentation would accelerate this goal and increase enterprise-user confidence.
-
Shared CI resources for cloud-provider end-to-end tests: The project integrates with multiple cloud backends (AWS Bedrock for embeddings, AWS Agent Registry / Amazon Bedrock AgentCore for federation, Amazon DocumentDB for storage, and optionally Microsoft Entra ID, Okta, Auth0, Google IdP for authentication). End-to-end tests requiring cloud provider credentials currently run locally on maintainer machines rather than in CI. Shared CI infrastructure with scoped credentials for multi-cloud e2e testing would let us expand CI coverage and prevent regressions on cross-cloud integration paths.
Additional Information
What makes this project distinctive. Most MCP and agent infrastructure projects in the market today solve one layer of the problem: a gateway, a registry, a security scanner, or a policy engine. MCP Gateway & Registry is the only open-source project we are aware of that unifies all four layers (gateway + registry + integrated security scanning + policy/auth) in a single, production-deployed control plane for both MCP servers and A2A agents. It is not a standards proposal or a reference implementation of a single spec; it is working software with 349 unique telemetry-observed deployments to date, 15 deployments confirmed actively running for 14+ days, and an ongoing new-install rate of roughly 20 new unique deployments per day. The project treats MCP and A2A as adopted standards to be governed in production, rather than as surfaces to be re-designed.
Why AAIF is the right home. The project already plays the role AAIF is being built to formalize: it is the neutral meeting point for multiple agentic AI standards (MCP, A2A, Anthropic Registry REST API, AWS Agent Registry federation). Today it lives under a community-spirited but AWS-led GitHub organization, which limits the trust enterprise adopters and other cloud providers can place in it. AAIF hosting would convert the project from "AWS's open-source contribution to agentic AI" into "the vendor-neutral governance and discovery layer for agentic AI," which is what the architecture was built to be. The project has already implemented the compatibility surfaces (Anthropic Registry REST API, federated registries, five identity providers) that make this neutrality credible; AAIF hosting provides the governance credential.
Commitments on acceptance. The project commits to, within 6 months of AAIF acceptance: (1) create a formal GOVERNANCE.md, MAINTAINERS.md, and CODEOWNERS aligned with AAIF's governance framework; (2) expand core maintainership to include contributors from at least one non-AWS organization; (3) launch a public bi-weekly community meeting with published notes; (4) achieve an OpenSSF Best Practices badge; (5) complete integration discussions with complementary AAIF-candidate projects (Envoy AI Gateway, agentgateway, DAP, PIC) to document composable deployment patterns for AAIF users.
Supporting evidence and artifacts:
- AWS Workshop Studio: The project is featured in an official AWS Workshop, "Securing AI Agent Ecosystems with MCP Gateway and Registry" (https://catalog.us-east-1.prod.workshops.aws/workshops/0c3265a6-1a4a-467b-ae56-e4d019184b0e/en-US).
- Demo videos: The AWS Show & Tell presentation on YouTube, plus multiple recorded demos covering OAuth flows, dynamic tool discovery, agent skills, and virtual MCP servers (linked from the project README).
- Documentation: 30+ topic-level documentation files covering setup guides, architecture, design documents, API references, and operational guides.
- Test Suite: 701+ passing tests with unit, integration, and E2E coverage running on GitHub Actions CI.
- Multi-platform deployment: Proven deployment patterns for Docker Compose, Podman, AWS ECS Fargate (Terraform), AWS EKS (Helm), and macOS local development.
- Presentation materials: Slide deck available at here
Application contact name(s) and email(s)
Amit Arora, aroraai@amazon.com
Contributing or sponsoring entity signatory information
| Name |
Address |
Type (e.g., Delaware corporation) |
Signatory name and title |
Email address |
| Amazon Web Services, Inc. |
410 Terry Avenue North, Seattle, WA 98109, USA |
Delaware corporation |
David Nalley |
djnalley@amazon.com |
Project Name
MCP Gateway & Registry
Project Description
MCP Gateway & Registry is a unified platform for centralizing access to AI Assets, which today include three categories: MCP (Model Context Protocol) servers, AI agents, and agent skills. The platform is designed so additional AI asset types can be added over time without restructuring the governance model. It serves four core functions: (1) a unified MCP Server Gateway providing a single access point for multiple MCP servers, (2) an MCP Servers Registry for registering, discovering, and governing MCP server access, (3) an Agent Registry and A2A Communication Hub for agent registration, discovery, governance, and direct agent-to-agent communication through the A2A protocol, and (4) an Agent Skills Registry for registering, discovering, and governing reusable agent skills, with per-skill authentication credentials and integrated security scanning.
The project originated in May 2025 to solve a growing problem in enterprise AI development: as teams adopt AI coding assistants (VS Code, Cursor, Claude Code) and deploy autonomous AI agents, each developer and agent must independently configure connections to dozens of MCP servers, manage scattered credentials, and operate without centralized governance. This creates security risks, credential sprawl, zero visibility into tool usage, and no standard for agent-to-agent discovery.
The platform replaces this chaos with a single control plane. Developers, AI coding assistants (VS Code, Cursor, Claude Code, and others), and AI agents connect to one gateway and gain access to all approved MCP servers with enterprise authentication (Keycloak, Microsoft Entra ID, Okta, Auth0, Amazon Cognito), fine-grained access control at the tool and method level, comprehensive audit logging, and real-time observability. Key capabilities include dynamic tool discovery via semantic search, virtual MCP servers (tool aggregation from multiple backends), peer-to-peer registry federation, agent skills registry, MCP server security scanning (via Cisco AI Defense), A2A protocol support, and integration with external registries (Anthropic MCP Registry, AWS Agent Registry / Amazon Bedrock AgentCore).
As of release 1.23.0, the project has 642+ GitHub stars, 165+ forks, 37+ contributors, 700+ passing tests, and has been deployed on AWS ECS, EKS, Docker Compose, and Podman across multiple organizations. An AWS Workshop Studio lab ("Securing AI Agent Ecosystems with MCP Gateway and Registry") is available for hands-on learning.
Alignment with AAIF Mission
MCP Gateway & Registry directly advances the Agentic AI Foundation's mission of fostering open, interoperable, and well-governed agentic AI systems:
Open Standards Adoption: The project is a production gateway and registry layer built around two foundational agentic AI standards: the Model Context Protocol (MCP) and the Agent-to-Agent (A2A) protocol. For MCP, the project acts as a gateway (proxying traffic to registered MCP servers) and implements the Anthropic MCP Registry REST API v0.1 specification so any MCP-compliant client can consume servers from the registry without modification. For A2A, the project is a registry-only layer that hosts A2A agent cards and provides authenticated discovery, while agents continue to communicate peer-to-peer using the A2A protocol (the project does not proxy or reimplement the A2A wire protocol). The project does not define new standards; it adopts these existing ones and adds the governance and discovery layer around them.
Interoperability and Federation: The platform was designed to avoid registry silos. It supports three federation modes: (a) peer-to-peer federation between independent MCP Gateway & Registry instances (so an agent registered in one enterprise's deployment can be discovered by a partner running their own deployment), (b) imports of upstream servers from the Anthropic MCP Registry, and (c) federation with AWS Agent Registry / Amazon Bedrock AgentCore. This ensures agents and tools remain portable across vendors and across organizational boundaries.
Governance and Trust: The project provides the governance layer enterprise adoption of agentic AI requires: fine-grained access control at the tool and method level, compliance audit logging, integrated security scanning of MCP servers, A2A agents, and agent skills (via Cisco AI Defense scanners), centralized credential management, and support for five enterprise identity providers (Keycloak, Microsoft Entra ID, Okta, Auth0, Amazon Cognito). This directly supports responsible AI agent deployment in regulated industries.
Community-Driven Development: The project is Apache 2.0 licensed with 37+ contributors from multiple organizations, 21+ releases in 12 months, comprehensive documentation, and an AWS Workshop Studio lab used by participants worldwide to learn secure agentic AI deployment patterns.
Relation to Existing AAIF Projects
MCP Gateway & Registry is designed to complement and extend the three projects currently hosted by AAIF (as listed in the AAIF landscape):
Model Context Protocol (MCP): This is the most direct relationship. MCP Gateway & Registry is a production governance, discovery, and gateway layer for the MCP ecosystem. It implements the Anthropic MCP Registry REST API specification so that any MCP-compliant client can consume servers from the registry without modification, imports servers from the upstream Anthropic MCP Registry, and applies enterprise policy (authentication, fine-grained authorization, audit logging, security scanning via Cisco AI Defense) before MCP servers are exposed to developers or AI agents. Where MCP defines the protocol, this project supplies the scaled, governed deployment layer that enterprise MCP adopters need.
goose: goose is a client-side extensible AI agent that connects to MCP servers. MCP Gateway & Registry is directly useful to goose users and operators: instead of each goose instance maintaining its own list of MCP server connections and credentials, goose can point at a single MCP Gateway endpoint and inherit the registry's governed catalog, authentication, and tool-level authorization. An initial goose integration already exists: the registry UI exposes goose as a first-class connection option alongside VS Code, Cursor, and Claude Code (see merged PR #1018), so goose users can connect to a governed MCP Gateway with a single click. We see goose + MCP Gateway & Registry as a natural "agent runtime + agent infrastructure" pairing for AAIF users, with room to deepen the integration over time.
AGENTS.md: AGENTS.md provides a standardized, predictable file format for giving context and instructions to AI coding agents inside a project. This project and AGENTS.md are complementary at different layers: AGENTS.md scopes per-repository agent instructions; MCP Gateway & Registry scopes organization-wide tool discovery, identity, and governance. A coding agent following AGENTS.md conventions can discover and invoke the right MCP tools (and A2A agents) through an AAIF-governed registry rather than hardcoded configuration.
Open standards the project implements (not AAIF-hosted, included here for context):
Agent-to-Agent (A2A) Protocol: The project provides a production-grade registry and discovery layer for A2A-capable agents. A2A defines the communication protocol; MCP Gateway & Registry solves the discovery and governance problem: how agents find each other, authenticate, and operate within organizational policies. Agents register A2A agent cards in the registry and discover other agents via semantic search, then communicate directly peer-to-peer using A2A. Supporting A2A alongside MCP is intentional: it keeps the governance layer neutral across the two dominant agentic interoperability standards.
Anthropic MCP Registry REST API: The project implements this upstream specification so that the MCP ecosystem's existing client and tooling base works against this registry unchanged.
Complementary candidate projects currently under AAIF review. Several projects with active proposals on
aaif/project-proposalsare natural architectural complements rather than competitors. Envoy AI Gateway (proposal #18) operates at the network/proxy layer for LLM and MCP traffic (token-based rate limiting, multi-provider routing, upstream authentication); MCP Gateway & Registry operates at the discovery, catalog, and identity-aware governance layer. A production deployment could run both: Envoy AI Gateway as the data-plane proxy, MCP Gateway & Registry as the control-plane catalog. Similarly, agentgateway (proposal #11) and policy-focused proposals like DAP (#17) and PIC Standard (#16) address overlapping governance concerns at adjacent layers. We would welcome joint architectural discussions under an AAIF umbrella if any of these are accepted.Example Use Cases and Evidence of Adoption
Use Case 1: Enterprise AI Coding Assistant Governance
Organizations with hundreds of developers using AI coding assistants (VS Code with Copilot, Cursor, Claude Code) deploy MCP Gateway & Registry to provide a single, governed configuration point. Instead of each developer managing their own MCP server connections and credentials, central IT manages a curated catalog of approved MCP servers accessible through the gateway with SSO authentication and fine-grained permissions.
Use Case 2: Autonomous Agent Ecosystems
Teams building multi-agent systems use the registry as a discovery and governance layer. Agents register their capabilities and discover other agents through semantic search, then communicate directly via the A2A protocol. The gateway provides M2M authentication (OAuth2 Client Credentials), access control, and audit trails for all agent interactions.
Use Case 3: AWS Workshop Studio Deployment
The project is featured in an official AWS Workshop Studio lab ("Securing AI Agent Ecosystems with MCP Gateway and Registry") where participants deploy the full stack on AWS, configure enterprise authentication, implement access control, and secure AI agent communications. This demonstrates production-readiness and educational value.
Evidence of Production Deployments:
Telemetry-Backed Adoption Data (as of 2026-05-11):
The project collects anonymous, opt-out telemetry (no PII, no hostnames). The following metrics are from the collection window ending 2026-05-11:
Community Signals:
public.ecr.aws/p3v1o3c6/) with regular releasesTechnical Committee Sponsor (if identified)
James Ward, Principal Developer Advocate, AWS.
GitHub Repository URL
https://github.com/agentic-community/mcp-gateway-registry
License
Apache 2.0
Governance Model
The project currently follows a maintainer-led governance model:
docs/design/directory with design documents.docs/design/and regular release notes inrelease-notes/. The project commits to launching a public bi-weekly community meeting with open attendance and published meeting notes within 30 days of AAIF acceptance, to provide a synchronous venue for roadmap discussion, contributor onboarding, and cross-project coordination with other AAIF projects.The project is prepared to adopt LF Minimum Viable Governance or the AAIF governance framework upon acceptance as a hosted project.
CI/CD & Release Workflow
The project uses GitHub Actions for CI/CD with 12 workflow files covering:
registry-test.ymlruns the full pytest suite (701+ tests) with 8 parallel workers on every PR targeting main/develop. Separate workflows test the auth server (auth-server-test.yml), Helm charts (helm-test.yml), Terraform configs (terraform-test.yml), and metrics service (metrics-service-test.yml).build-registry.yml,build-auth-server.yml,build-mcpgw.ymlbuild and push multi-architecture container images to Amazon ECR Public (public.ecr.aws/p3v1o3c6/).release-images.ymlbuilds multi-architecture (AMD64/ARM64) Docker images and pushes tagged releases. Helm chart updates are automated viahelm-chart-update.ymlandhelm-release-retag.yml.docs.ymlbuilds and deploys documentation to GitHub Pages.Release Frequency: The project has shipped 21 releases since May 2025 (roughly biweekly to monthly), with detailed release notes in the
release-notes/directory. Each release follows semantic versioning.Public-Facing Contribution Process for Specifications
The project's contribution process for specifications and design changes:
Design Documents: Major features are proposed and documented in
docs/design/with comprehensive design documents covering architecture, data models, API contracts, and security considerations. Examples include federation architecture, virtual MCP server design, hybrid search architecture, authentication design, and AWS Agent Registry federation.GitHub Issues: All feature proposals start as GitHub Issues with detailed descriptions, acceptance criteria, and community discussion. The project uses milestones to group issues into releases.
Pull Request Review: All pull requests are reviewed and tested by one of the core maintainers before merge. Review guidelines (including test-pass requirements and a "Merge Specialist" persona for comprehensive PR evaluation) are documented in the project's
CLAUDE.md.Contributing Guide: CONTRIBUTING.md documents the full contribution workflow including issue filing, PR submission, and code review expectations.
Publicly Accessible Issue Tracker
https://github.com/agentic-community/mcp-gateway-registry/issues
Currently 100+ open issues with active community engagement. Across the project's ~12 month history, the repository has accumulated 400+ issues (101 open, 299 closed) and 590+ pull requests (20 open, 500+ merged), for a combined total of nearly 1,000 tracked changes. Issues are organized into per-release and per-week milestones (recent releases v1.0.20, v1.0.21, v1.0.22, and 1.23.0 are all complete; 1.24.0 and 1.25.0 are in progress) with labels for categorization. The project also uses GitHub Discussions for feature requests and general discussion.
External Project Dependencies
Key runtime dependencies (all open-source):
Infrastructure dependencies:
Maintainers & Contributors
Core Maintainers:
Note: The project has two AWS-employed core maintainers today. The broader contributor base spans 37+ individuals across multiple organizations. Expanding to multi-organization core maintainership is a committed post-acceptance goal with a 6-month target, promoting active contributors based on sustained technical contributions and code review participation. AAIF hosting would provide the formal governance framework (documented maintainer promotion criteria, contribution-based advancement) to make this transition credible and repeatable.
Active Contributors (37+ total, top non-maintainer contributors by commit count as of 2026-05-11):
Bot contributors (Dependabot, GitHub Actions) are excluded from the table above. Full contributor list: https://github.com/agentic-community/mcp-gateway-registry/graphs/contributors
Leadership Team & Decision Process
Project Lead: Amit Arora (@aarora79) provides overall technical direction and coordinates the roadmap.
Decision-Making Process:
docs/design/before implementation.The project is open to adopting a more formal governance structure (such as a Technical Steering Committee) as part of joining AAIF.
Roadmap
The project maintains a public roadmap organized into release milestones on GitHub:
Completed (recent):
cloud=unknowntelemetry bucket), IAM hardening (M2M orphan delete + deduplicated user listing), Splunk-ready JSON-Lines logging, and a switch to strict semantic versioning (novprefix) for all tags, images, and Helm chart versions.Planned (next 3 months):
/metricsendpoint for the registry, A2A reverse-proxy gateway support, dependency management for agent/service/skill cards, Registry Copilot (embedded AI chat assistant for registry operations), Helm ALB ingress-scheme configurability, UI title configurability, and MCP registration deduplication.WWW-Authenticateflow, Entra v1 scope verbatim support, RFC 8707 enforcement, and a documented no-DCR decision.Backlog (6-12 months):
Full roadmap with links to all issues:
Security
The project does not currently hold an OpenSSF Best Practices badge. However, the project demonstrates strong security practices:
Pursuing an OpenSSF Best Practices badge is a near-term goal that AAIF hosting would help accelerate.
Website URL
https://agentic-community.github.io/mcp-gateway-registry/
Documented Governance Practices (if any)
No formal
GOVERNANCE.md,MAINTAINERS.md, orCODEOWNERSfile exists today. The project commits to creating all three aligned with AAIF's governance framework within 6 months of acceptance. DCO (Developer Certificate of Origin) sign-off will be adopted on all commits as part of this transition, replacing the project's current Amazon Open Source Code of Conduct attestation flow with the standard AAIF/LF contribution process.Links to Social Media Accounts
The project does not yet maintain dedicated social accounts. Project updates, new release announcements, and community posts are currently shared every 2 to 3 weeks via the project lead's personal LinkedIn account: Amit Arora. An example of a recent release post is here. On acceptance into AAIF, this will be formalized with a dedicated project LinkedIn page (and an X account) within 30 days, so that community engagement and release communications live under the project's own identity rather than an individual contributor's.
Trademark and accounts
Details of Existing Financial Sponsorship
The project receives engineering contributions from AWS, which employs the two current core maintainers (Amit Arora, Omri Shiv) and funds their contributions to the project. Pre-built multi-architecture (AMD64/ARM64) container images are published to Amazon ECR Public under
public.ecr.aws/p3v1o3c6/and referenced by the project's Helm charts. GitHub Actions provides CI/CD via GitHub-hosted runners. Documentation is hosted on GitHub Pages.Infrastructure Needs or Requests
The project's current infrastructure is largely self-hosted or funded by AWS, but two AAIF-provided resources would materially strengthen the project:
Security audit tooling and OpenSSF Best Practices badging support: The project is committed to achieving an OpenSSF Best Practices badge (see Security section). Foundation-facilitated access to third-party security audit tooling, SBOM scanning infrastructure, and AAIF-community review of our security posture documentation would accelerate this goal and increase enterprise-user confidence.
Shared CI resources for cloud-provider end-to-end tests: The project integrates with multiple cloud backends (AWS Bedrock for embeddings, AWS Agent Registry / Amazon Bedrock AgentCore for federation, Amazon DocumentDB for storage, and optionally Microsoft Entra ID, Okta, Auth0, Google IdP for authentication). End-to-end tests requiring cloud provider credentials currently run locally on maintainer machines rather than in CI. Shared CI infrastructure with scoped credentials for multi-cloud e2e testing would let us expand CI coverage and prevent regressions on cross-cloud integration paths.
Additional Information
What makes this project distinctive. Most MCP and agent infrastructure projects in the market today solve one layer of the problem: a gateway, a registry, a security scanner, or a policy engine. MCP Gateway & Registry is the only open-source project we are aware of that unifies all four layers (gateway + registry + integrated security scanning + policy/auth) in a single, production-deployed control plane for both MCP servers and A2A agents. It is not a standards proposal or a reference implementation of a single spec; it is working software with 349 unique telemetry-observed deployments to date, 15 deployments confirmed actively running for 14+ days, and an ongoing new-install rate of roughly 20 new unique deployments per day. The project treats MCP and A2A as adopted standards to be governed in production, rather than as surfaces to be re-designed.
Why AAIF is the right home. The project already plays the role AAIF is being built to formalize: it is the neutral meeting point for multiple agentic AI standards (MCP, A2A, Anthropic Registry REST API, AWS Agent Registry federation). Today it lives under a community-spirited but AWS-led GitHub organization, which limits the trust enterprise adopters and other cloud providers can place in it. AAIF hosting would convert the project from "AWS's open-source contribution to agentic AI" into "the vendor-neutral governance and discovery layer for agentic AI," which is what the architecture was built to be. The project has already implemented the compatibility surfaces (Anthropic Registry REST API, federated registries, five identity providers) that make this neutrality credible; AAIF hosting provides the governance credential.
Commitments on acceptance. The project commits to, within 6 months of AAIF acceptance: (1) create a formal
GOVERNANCE.md,MAINTAINERS.md, andCODEOWNERSaligned with AAIF's governance framework; (2) expand core maintainership to include contributors from at least one non-AWS organization; (3) launch a public bi-weekly community meeting with published notes; (4) achieve an OpenSSF Best Practices badge; (5) complete integration discussions with complementary AAIF-candidate projects (Envoy AI Gateway, agentgateway, DAP, PIC) to document composable deployment patterns for AAIF users.Supporting evidence and artifacts:
Application contact name(s) and email(s)
Amit Arora, aroraai@amazon.com
Contributing or sponsoring entity signatory information