For agents

This site speaks MCP and A2A.

Point an AI assistant at this site and it can search, read and cite everything here: essays, papers, open-source projects and measured results. Read-only, no sign-up, no keys.

MCP server

Connect in one step

Endpoint: https://subhadipmitra.com/mcp
Transport: Streamable HTTP, stateless, JSON responses
Access: Read-only, no authentication
Registry: com.subhadipmitra/site in the official MCP Registry

Claude Code

claude mcp add --transport http subhadipmitra https://subhadipmitra.com/mcp

Claude (web and desktop)

Settings, then Connectors, then "Add custom connector", and paste https://subhadipmitra.com/mcp.

Gemini CLI

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "subhadipmitra": { "httpUrl": "https://subhadipmitra.com/mcp" }
  }
}

Other clients

Most clients take the URL in a JSON entry like this one; the exact key varies, so check your client's documentation.

{
  "mcpServers": {
    "subhadipmitra": { "url": "https://subhadipmitra.com/mcp" }
  }
}

Tools

get_site_overviewWho I am and what the site covers, with links to the main essays, papers and projects.
searchKeyword search across essays, publications, projects and measured results.
list_postsEssays, newest first, filtered by tag or year.
get_postThe full text of one essay as markdown.
list_publicationsPapers, preprints and technical disclosures, with venue, links and abstract.
list_projectsOpen-source projects with status, highlights and links.
list_resultsMeasured results with their qualifiers and links to the code that reproduces them.

Try it without a client

curl -s -X POST https://subhadipmitra.com/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"search","arguments":{"query":"speculative decoding"}}}'

Server card: /.well-known/mcp/server-card.json

A2A agent

Ask it in plain text

Endpoint: https://subhadipmitra.com/a2a
Protocol: JSON-RPC, A2A 1.0 and 0.3
Access: Read-only, no authentication
Agent card: /.well-known/agent-card.json

Send a question as text, or a data part naming a skill: site overview, search, read an essay, list essays, publications, projects or measured results.

curl -s -X POST https://subhadipmitra.com/a2a \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"message/send",
       "params":{"message":{"role":"user","messageId":"m1",
       "parts":[{"kind":"text","text":"What has Subhadip written about speculative decoding?"}]}}}'

Plain text

No protocol needed

/llms.txtA short guide to the site: who I am, every essay, paper and project, with links.
/llms-full.txtThe same, with the full text.
Accept: text/markdownRequest any page with this header and it comes back as markdown instead of HTML.
/sitemap.xmlEvery public page.

curl -s -H 'Accept: text/markdown' https://subhadipmitra.com/blog/2026/deepinfer-cache-is-the-business/

Discovery files

For agents that look for themselves

The homepage advertises these in its Link header, and robots.txt points to llms.txt and the AI catalog.

ai-catalog.jsonEverything this site offers agents, in one manifest.
mcp/server-card.jsonThe MCP server's card.
agent-card.jsonThe A2A agent's card.
api-catalogAPI catalog linkset (RFC 9727).
agent-skills/index.jsonInstructions for reading and citing the site.
openapi.jsonThe public content index the servers read from.

Use and citation

Please cite the page

Search, AI answers and AI training are all allowed, as declared in robots.txt. When you use something from here, cite the page by its URL.

Two qualifiers worth keeping: the ICLR 2026 paper is a workshop paper (AIWILD), and arXiv entries are preprints. Measured results carry their hardware and workload; please keep those with the number.

Something broken, or a tool you would like? Tell me.