Support for A2A AuthN and AuthZ

[dmiyamasu]

MCP and A2A offer solutions at different layers of the agentic AI stack, and both may exist separately from each other.

One area that A2A excels is enterprise authentication and authorization features

MCP is making an effort to offer this in a client-server context, but A2A is thinking about infinite agent scale. i.e. N agents calling N agents using N Oauth providers.

I'd love to see more robust enterprise features in mcp-context-forge - maybe add this one to the roadmap?

NOTE: Application to Application partial support does not appear to be the same thing as agent to agent authn and authz

[crivetimihai]

Hi there, thanks for raising this! You're spot-on: our goal is a first-class, enterprise-grade MCP Gateway with AuthN/AuthZ between agents and tools, Kubernetes-native deployment with a Horizontal Pod Autoscaler, stateless replicas via Redis, Alembic-managed database migrations, configurable gunicorn and FastAPI runtime, secure minimal container images with non-root privileges and observability through built-in metrics and tracing.

We're also focused on a good developer experience to drive adoption, so pip install mcp-contextforge-gateway && mcpgateway to get started should be sufficient.

Where it fits on the roadmap

Authentication, Authorization, Security & Identity

• 🔜 JWT Token Catalog – short-lived, revocable tokens generated via API
• 🔜 Per-Virtual-Server API Keys – credentials scoped to one MCP server
• 🔜 Role-Based Access Control – user, team, and global scopes

Federation & Routing

• 🟡 A2A-inspired transport – partial today; full duplex next
• 🟡 OpenTelemetry – basic metrics live, distributed tracing coming

AuthN/AuthZ/RBAC

• 🔜 OAuth2 / OIDC bearer tokens (lands with RBAC)
• 🔜 Fine-grained scope enforcement via RBAC (looking at private, team, public scopes for the catalogs)
• 🔜 Tool annotations (Include tool annotations #176) for read-only, destructive, idempotent, or external-access calls

Deployment

We've added detailed deployment documentation and scripts, and improving our CI/CD pipeline https://ibm.github.io/mcp-context-forge/deployment/

• ✅ TLS-only endpoints fully supported, see make docker-run-ssl-host for local test sample
• ✅ Images based on latest ubi9-micro with latest updates, using build stages and dropping to a regular user to reduce footprint
• ✅ Helm chart with startup / readiness / liveness probes, resource limits, and an HPA
• ✅ Redis for cross-pod session and task state
• ✅ SQLAlchemy (PostgreSQL or other) with Alembic migrations
• ✅ Stateless API nodes—safe to scale horizontally

Quality & security automation

• ✅ CI runs unit and integration tests with pytest, lint, bandit, etc. on every PR
• ✅ Ruff + mypy enforce style and type safety; Black auto-formats
• ✅ Trivy scans container images and dependencies for CVEs
• ✅ Automatic deployment to cloud as part of CICD
• 🔜 Additional automatic deployment that includes Alembic migrations, and deployment with PostgreSQL and Redis across AWS, Azure, ArgoCD (Kubernetes) coming soon as part of GitHub actions.
• 🟡 Integrate SonarQube into GitHub Actions (currently there's a local make sonar-submit-docker for SonarQube scans)

How you can help

Open a Feature issue using the provided template, describing additional security, enterprise features, etc. or submit a PR - would love to hear more from you if you have other ideas or features.