Service
Auth for AI Agents & MCP
AI agents need identity too. OAuth 2.0 for MCP servers, scoped credentials for autonomous processes, passkey-embedded agent flows, and audit trails that hold up under scrutiny.
Identity for Autonomous Systems
The agentic era is arriving faster than identity infrastructure is adapting. AI agents — whether running Claude via MCP, GPT via function calling, or custom autonomous processes — are accessing APIs, reading sensitive data, and taking real actions on behalf of users.
Most deployments today handle agent auth as an afterthought: a shared API key, a hardcoded credential, or the user's own session token passed through. DVLY builds identity infrastructure that treats agents as first-class principals.
The MCP Authorization Problem
The MCP Authorization specification (2025-03-26) mandates OAuth 2.1 with PKCE for MCP servers accessing protected resources. Without a proper authorization server, your MCP server is either insecure (open) or inaccessible (blocked by enterprise security teams). DVLY implements the full spec.
OAuth 2.0 for MCP Servers
Full MCP Authorization spec implementation: dynamic client registration, authorization code + PKCE, token introspection, protected resource metadata endpoint, and consent screens.
Agent Identity
Dedicated identity principals for autonomous agents — separate from the users who authorize them. Agents get their own client credentials, bound to specific tasks and time windows.
Scoped Credentials
Short-lived, least-privilege tokens scoped to exactly what an agent needs: read-only access to one dataset, write access to one endpoint, delegated on behalf of a specific user.
Passkey-Embedded Agent Flows
WebAuthn-based human-in-the-loop authorization for high-privilege agent actions — the agent requests authorization, the user approves with a passkey tap, no password required.
Attribution & Audit Trails
Every agent action is attributed to both the agent identity and the authorizing human. Structured logs feed into your SIEM or compliance tooling.
Zero-Trust Agent Policies
Policy enforcement at the authorization server level: agents can't accumulate permissions across sessions, can't self-escalate, and can't operate outside their authorized scope.
Common Agentic Auth Patterns
Delegated Authorization
A user authorizes an AI assistant to act on their behalf within defined limits. The agent receives a delegated token with the user's identity claims but constrained scopes. When the task is complete or the time window expires, the token is revoked.
Service-to-Service Agent Auth
Fully automated pipelines where no human is in the loop at runtime. Agents authenticate using client credentials tied to specific pipeline definitions — not shared API keys. Each pipeline has its own identity, rotation schedule, and audit log.
Human-in-the-Loop Escalation
Agents operating within pre-approved scopes can proceed autonomously. When an action requires elevated permissions, the agent pauses and presents a passkey-authenticated consent request to the authorizing human before proceeding.
Multi-Agent Orchestration
Complex agentic systems involve multiple agents delegating to each other. DVLY implements trust hierarchies so that downstream agents can't exceed the permissions of the orchestrating agent — preventing privilege escalation through chaining.
Frequently Asked Questions
What is MCP and why does it need OAuth?
MCP (Model Context Protocol) is Anthropic's open protocol for connecting AI assistants to external data sources and tools. The MCP Authorization specification (2025-03-26) mandates OAuth 2.1 with PKCE for any MCP server that grants access to protected resources. Without proper OAuth flows, MCP servers are either insecure or inaccessible to enterprise deployments.
How does DVLY define "agent identity"?
Agent identity means giving an autonomous AI process its own cryptographic identity — separate from the human user who authorized it. This allows you to scope what the agent can do, attribute actions to the agent specifically, and revoke access independently of the user's session.
Can DVLY implement OAuth 2.0 for our existing MCP server?
Yes. DVLY can implement the full MCP authorization server spec including dynamic client registration, authorization code flow with PKCE, token introspection, and the protected resource metadata endpoint. We work with Auth0, Okta, and custom authorization servers.
What are scoped credentials for AI agents?
Scoped credentials are short-lived access tokens or certificates that grant an agent only the specific permissions it needs for a given task — read-only access to one database, write access to one API endpoint, etc. This enforces least-privilege for autonomous processes that may otherwise accumulate excessive access.
What does an audit trail for AI agents include?
DVLY's agent audit trails record the full attribution chain: which human authorized the agent, which agent identity acted, what resources were accessed, what data was read or written, and when credentials were issued and expired. This is essential for compliance, debugging, and security incident response.
Building with AI agents?
Let's talk about how to give them proper identity before you ship to production.
Talk to DVLY