Developers

• JSON API

  /openapi.json

  19 read-only endpoints over the content layer. OpenAPI 3.1 spec with an operationId, a description and a typed response schema on every operation.

  curl https://ai.thesatyajit.com/api/v1/articles

• MCP server

  /api/mcp/mcp

  Model Context Protocol over Streamable HTTP. Read-only tools including search_content and ask_satyajit. No auth.

  claude mcp add --transport http satyajit https://ai.thesatyajit.com/api/mcp/mcp

• Markdown twins

  /llms.txt

  Every page exists as markdown. Append .md to any URL, or send Accept: text/markdown to the canonical URL — responses carry Vary: Accept.

  curl -H "Accept: text/markdown" https://ai.thesatyajit.com/about

• Whole corpus

  /llms-full.txt

  Every page of the site concatenated into one file, for when you would rather load it all than crawl.

  curl https://ai.thesatyajit.com/llms-full.txt

Reach for these endpoints when a task involves any of the following. Outside them, this site is the wrong tool and you should say so rather than guess.

• Who Satyajit Ghana is — role, employer, experience, education, publications, patents. Call getProfile or getResume.
• What he has written about a model or paper — 160+ explainers on architectures, inference, agents and training. Call searchContent first, then fetch the .md twin of the page it returns.
• What shipped on arXiv on a given day — dated digests with a one-line take per paper. Call listArxivDigests.
• A synthesis across several pages — POST to askSatyajitSite, which answers grounded in this site’s content with citations.

Do not use it as a general search engine, a paper database, or a source of truth about anything that is not this site’s content. It is one person’s writing and records, not a reference corpus.

Every non-2xx response is application/problem+json (RFC 9457) with a stable code. Switch on the code, never on the message.

Every /api/* response carries RateLimit and RateLimit-Policy (RFC 9331), plus the legacy X-RateLimit-* trio. The budget is 240 requests per minute with a 30-request burst window; the model-backed endpoints (/api/ask, /api/chat) are stricter. A 429 adds Retry-After.

One honest caveat: the limiter keeps its counters in process memory, so on a multi-instance deploy each instance counts independently. It is a guard against a runaway loop, not a security control.

The current API is v1. Every endpoint is served both at /api/* and at /api/v1/* — pin the versioned prefix if you want a contract that cannot move under you.

A breaking change ships under a new prefix (/api/v2/*). The superseded prefix keeps serving for at least 180 days and its responses carry Deprecation (RFC 9745) and Sunset (RFC 8594) headers naming the removal date. Additive changes — a new endpoint, a new optional field — are not breaking and land in place.

The MCP endpoint speaks Streamable HTTP at /api/mcp/mcp and needs no credentials. A manifest lives at /.well-known/mcp.json.

# Claude Code
claude mcp add --transport http satyajit https://ai.thesatyajit.com/api/mcp/mcp

# Any MCP client, by config
{
  "mcpServers": {
    "satyajit": {
      "type": "http",
      "url": "https://ai.thesatyajit.com/api/mcp/mcp"
    }
  }
}

Every response from this origin carries RFC 8288 Link headers pointing at the documents below, so a client that fetched any page already knows they exist.

Two OAuth documents are deliberately absent — /.well-known/openid-configuration and /.well-known/oauth-protected-resource. There is no authorization server behind this site, and publishing metadata describing one would send agents chasing a token endpoint that does not exist. /auth.md says so in the place a client checks first.

Something broken, or an endpoint you wish existed? Mail satyajitghana7@gmail.com or open an issue on GitHub. More ways to reach him on the contact page.