Den Delimarsky
Conference affiliation: Microsoft · 2025
Den Delimarsky is a Member of Technical Staff at Anthropic and a lead maintainer of Model Context Protocol, the open standard connecting AI applications to tools and data. His authorization work addresses a practical problem behind that connection: letting an assistant use protected systems while preserving the permissions of the person who asked it to act.
On this page
From developer tools to protocol leadership
After six years at Microsoft, Den joined Amazon Web Services in 2020 as a Senior Product Manager, Technical, working on Amazon EventBridge and event-driven infrastructure. He returned to Microsoft before joining Netlify in 2021 as a Group Product Manager focused on developer tools and web experiences.
During his third period at Microsoft, Den worked in Microsoft Security before moving into Developer Division and CoreAI. As a Principal Product Engineer, his work included authentication, authorization, developer tooling, and GitHub Copilot adoption. He served on the MCP steering committee and became a core maintainer focused on authorization and security. He also helped launch GitHub Spec Kit, crediting John Lam’s research in his account of leaving Microsoft.
Den joined Anthropic in January 2026 and was appointed an MCP lead maintainer that April. His responsibilities span the specification, SDKs, security, governance, and developer experience. His independent projects include PowerToys Awake, DeckSurf, and BlogScroll.
Protecting tools without building a token service
Den helped develop an MCP authorization model that separates two jobs: providing a tool and issuing the credentials needed to use it. The March 2025 specification required protected MCP servers to take on authorization-server responsibilities, including issuing and managing tokens. Developers building an integration therefore had to implement security infrastructure they might otherwise obtain from an established identity provider.
In his joint session with Julia Kasper on protected MCP servers, Den explained the replacement architecture. A protected MCP server responds to an unauthenticated request with an HTTP 401 and a pointer to protected-resource metadata. That metadata identifies the authorization service. The client discovers the service, completes the OAuth flow, and returns with a token; the MCP server validates it. Established identity providers handle token issuance and lifetime management. OAuth resource indicators bind tokens to their intended resources.
This division reduces the security code a tool developer must write without removing the server’s responsibility to check access. Den demonstrated it with the C# SDK and standard ASP.NET Core authentication and token-validation components. Julia demonstrated the separate Azure API Management and VS Code integrations. The shared architecture lets a protected API retain user-specific permissions rather than treating every assistant request as having the same access.
Den’s writing on authentication brokers develops a related concern: making existing identity infrastructure usable from developer tools. His MCP demonstration makes that concern concrete by relying on standard OAuth libraries and framework-supported validation instead of asking every server author to become an authorization-server implementer.
Specifications, hostile inputs, and interactive tools
Den’s other work addresses distinct decisions developers face when giving AI software more capabilities:
- Make requirements reviewable before implementation. GitHub Spec Kit supports specification-driven development through written requirements, technical plans, and implementation tasks. In Den’s description of the workflow, developers validate these stages before proceeding. The intermediate artifacts give people a chance to catch a misunderstood requirement before an agent builds code around it.
- Treat retrieved material as a possible source of hostile instructions. With Sarah Young, Den examined indirect prompt injection through documents, webpages, emails, and manipulated MCP tool descriptions. An assistant can encounter an attack while carrying out a legitimate request: the content it retrieves may try to redirect its behavior. Their guidance combines separation of trusted instructions from external data with approved dependencies, monitoring, and software-supply-chain safeguards.
- Give interactive tool interfaces an isolated place to run. Den’s writing on MCP Apps describes tools that return interactive HTML interfaces using existing protocol abstractions. Those interfaces run inside host-controlled, sandboxed frames. The tool can offer interaction beyond a text response while the host controls how the interface executes.
Read the topics behind these talks
1 conference talk
Loading more talks and highlights…
