MCP defines how AI agents connect to tools. It says nothing about how agents discover which tools exist. Without a discovery layer: n agents × m servers = n×m manual configuration decisions. With a registry: n+m.
The proposal: publish a single _mcp TXT record at your domain. Any compliant agent resolves it and finds your entire MCP ecosystem — no central directory, no SDK, no new protocol. The registry is itself an MCP server, so discovery uses the same tools/list call agents already make. Zero new client behaviour required.
You can verify the live DNS record right now:
dig TXT _mcp.mariothomas.com +short
Three servers registered: articles and locations (public), documents (authenticated — demonstrates the governance pattern). All endpoints speak JSON-RPC 2.0 over HTTPS.
Architecture paper, full spec, and CloudFormation template: github.com/mariothomas/mcp-dns-registry
Deployed on Lambda@Edge and DynamoDB. Costs under $5/month to run. Took one day to deploy.
A discussion is open in the MCP GitHub repository if anyone wants to engage on the proposal formally.
MCP defines how AI agents connect to tools. It says nothing about how agents discover which tools exist. Without a discovery layer: n agents × m servers = n×m manual configuration decisions. With a registry: n+m.
The proposal: publish a single _mcp TXT record at your domain. Any compliant agent resolves it and finds your entire MCP ecosystem — no central directory, no SDK, no new protocol. The registry is itself an MCP server, so discovery uses the same tools/list call agents already make. Zero new client behaviour required.
You can verify the live DNS record right now: dig TXT _mcp.mariothomas.com +short
Three servers registered: articles and locations (public), documents (authenticated — demonstrates the governance pattern). All endpoints speak JSON-RPC 2.0 over HTTPS.
Architecture paper, full spec, and CloudFormation template: github.com/mariothomas/mcp-dns-registry Deployed on Lambda@Edge and DynamoDB. Costs under $5/month to run. Took one day to deploy. A discussion is open in the MCP GitHub repository if anyone wants to engage on the proposal formally.