Skip to main content

Whitepapers

MCP Registry Architecture

A deployable, serverless architecture for MCP agent discovery: a DNS _mcp TXT record convention that points any compliant agent at your organisation's registry.

Version 1.5 Revised 19 March 2026

Overview

This paper proposes a lightweight, deployable architecture for solving the agent discovery problem in Model Context Protocol (MCP) ecosystems. The core proposal is a DNS-based convention — an _mcp TXT record — that points any compliant AI agent to an organisation’s MCP registry. DNS-based discovery is not new: MX, SRV, _dmarc, and WebFinger all use the same pattern. The novelty here is the specific convention for MCP, the design decision to make the registry itself an MCP server (requiring no special client code), and a fully serverless reference implementation.

No novel infrastructure is required. The proposal composes existing, operationally proven technologies into a governance-first discovery layer that any organisation can deploy in a day and run for under $5 per month at typical registry volumes.

What the paper covers

  • The agent discovery problem and why manual configuration does not scale
  • The _mcp DNS TXT record convention and its field specification
  • A fully serverless registry architecture: Amazon CloudFront, Lambda@Edge, and DynamoDB Global Tables
  • Public and private server access control enforced at the edge
  • Security considerations including DNSSEC, JWT authentication, and prompt injection via registry
  • A complete implementation guide with deployment steps and cost analysis
  • Relationship to the official MCP registry, SEP #1959, NANDA, A2A, and WebFinger
  • A live reference implementation at mcp.mariothomas.com

Reference implementation

A working implementation is live at mcp.mariothomas.com, including a queryable _mcp DNS record, a fully functional registry, and three MCP servers — public articles, public locations, and private documents with signed-URL retrieval. Source code is available in the GitHub repository.

Downloads

Markdown & Word editions

Alternate formats, the specification, and the reference implementation live in the companion GitHub repository.

Open the repository

History

Provenance and revision history, so what you are reading, and which version of it, is never in doubt.

Paper
MCP Registry Architecture
Status
Published
Version
1.5
Origin
First published 25 February 2026 as a discussion draft; the reference implementation went live at mcp.mariothomas.com in March 2026.
Repository
github.com/mariothomas/mcp-dns-registry

Versions

  1. 1.5 Expanded Section 2.3 with the explicit 255-byte DNS TXT record constraint as the hard technical rationale for the pointer-not-container design. Added the version field to the Section 5.4 DynamoDB schema, Section 5.7 — The Read Path: Audit Trail — and Section 7.5 — Context Window Constraints in Large Registries. Acknowledgements updated for Cole Johnston’s feedback.
  2. 1.4.1 Added copyright notice. No content changes.
  3. 1.4 Updated Section 8.1 to reflect that the reference implementation uses path-based routing; corrected Section 8.4/8.5 filenames and handler; rewrote Section 12 in present tense for the live deployment.
  4. 1.3 Extended Section 8.1 to document path-based and subdomain-based registry URL patterns as equally compliant approaches, with trade-offs. Updated SPEC.md accordingly.
  5. 1.2 Extended Section 7.4 on registry-level content filtering as a prompt-injection mitigation. Added Section 10.6 — Agent Peer Discovery: A Natural Extension. Acknowledgements section added.
  6. 1.1 Added Section 2.3 — What This Proposal Does Not Solve — clarifying that the _mcp record addresses discovery only; authentication, authorisation, and capability enumeration are out of scope for the DNS layer.
  7. 1.0 Initial publication.

I use cookies to understand how my website is used. This data is collected and processed directly by me, not shared with any third parties, and helps me improve my website. See my privacy and cookie policies for more details.