*The Problem:* As AI agents become more capable, they need secure ways to access tools and data. MCP (Anthropic's open protocol) is great for standardizing agent-to-tool communication, but deploying MCP servers in production is complex. You need authentication, rate limiting, audit logs, multi-tenancy, and more—all while ensuring your agents can't be exploited through prompt injection or other attacks.
*Our Solution:* One endpoint (agentictrust.com) that handles all your MCP servers. You write the tool logic, we handle everything else: - OAuth 2.0 authentication with scoped permissions - Rate limiting and usage analytics - Audit trails for compliance - Automatic versioning and routing - Protection against prompt injection attacks
*Technical Details:* We've also been working on OIDC-A (OpenID Connect for Agents), a proposal to extend OIDC for agent identity. It adds claims for agent attestation, delegation chains, and capabilities. This was recently featured by WorkOS's CEO at Identiverse.
The idea is that agents should have verifiable identities just like users do. When an agent acts on behalf of a user, you need to track that delegation chain for security and compliance.
*Why Now:* With Microsoft announcing MCP support in Windows 11 and OpenAI adopting the protocol, we're seeing explosive growth in MCP usage. But most implementations are insecure—exposed endpoints, no auth, vulnerable to attacks. We're fixing that.
*Links:* - Platform: https://agentictrust.com - OIDC-A Proposal: https://subramanya.ai/2025/04/28/oidc-a-proposal/ - WorkOS article on our work: https://workos.com/blog/identity-for-ai-agents
We're in early access and would love feedback from the HN community. What security concerns do you have about AI agents? How are you handling agent authentication today?