PyyanThe AI Index

Pyyan / MCP

The Pyyan MCP server

Point an agent at one URL and it can search the AI index, read sourced news, look up a model's specifications with the date each one was checked, and see what is trending. Free, unauthenticated, read only.

The endpoint

https://pyyan.com/api/mcp

Streamable HTTP transport, protocol 2025-06-18, with 2025-03-26 and 2024-11-05 accepted for older clients. Stateless: no session id is issued, so there is nothing to hold onto between calls and nothing to expire.

Connecting

Claude Code, one line:

claude mcp add --transport http pyyan https://pyyan.com/api/mcp

Claude Desktop, Cursor, Windsurf and anything else that takes a JSON config:

{
  "mcpServers": {
    "pyyan": {
      "type": "http",
      "url": "https://pyyan.com/api/mcp"
    }
  }
}

A client that only speaks stdio can bridge to it. No credentials are needed because there is nothing to authenticate:

{
  "mcpServers": {
    "pyyan": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://pyyan.com/api/mcp"]
    }
  }
}

Or speak to it directly. It is JSON-RPC over POST:

curl -X POST https://pyyan.com/api/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Tools

9 of them, all read only and all annotated as such, so a client that asks before running a tool can tell there is nothing here to approve.

search_ai_news

Search dated AI news from pyyan.com. Every item was read from a primary source before it was written up and carries the link back. Use this for what happened recently in AI: model releases, lawsuits, regulation, funding, chips and safety. Returns summaries; call get_ai_news_story for the full long form of one item.

  • query Free text, matched against headline, summary and organisations.
  • category One of Models, Law, Money, Policy, Safety, Silicon, Benchmarks, Open weights.
  • org An organisation named in the story, e.g. OpenAI or NVIDIA.
  • from ISO date, inclusive lower bound, e.g. 2026-09-01.
  • to ISO date, inclusive upper bound.
  • limit

get_ai_news_story

The full write up of one story by slug, including the long form where it has one: the hook, what happened, why it is different, a dated lineage of how the field got here, and what it does and does not mean. Get slugs from search_ai_news.

  • slug required The story slug.

search_ai_index

Search everything pyyan.com tracks: language models, open weight models, image, video, voice and OCR models, GPUs and AI chips, GPU clouds, agent tools and frameworks, routers, model hubs, vector databases. Every specification carries the source it came from, the date it was checked and a confidence level, so an answer built on this can be attributed. Use list_categories first if you are unsure which category to filter on.

  • query Free text, matched against name, maker and tagline.
  • category A category slug, e.g. language-models, open-models, gpu-chips.
  • maker Exact maker name, e.g. Anthropic, OpenAI, NVIDIA.
  • limit

get_ai_index_entry

One model, chip, cloud or tool by slug, with every specification and the source and date behind each one. Get slugs from search_ai_index.

  • slug required The entry slug, e.g. gpt-6-astra.

list_categories

The categories the index is organised into, each with what it ranks on and which specification fields its entries carry. Call this before search_ai_index when you do not already know the category slug.

search_people

Founders, researchers, safety people, pioneers, thinkers, educators and journalists in AI, with a short biography and what each is known for.

  • query Free text, matched against name, headline and biography.
  • tier Founders, Research, Safety, Pioneers, Thinkers, Educators, Journalists, Fiction.
  • limit

search_videos

A curated list of AI talks, interviews, lectures and podcasts, each with a line on why it is worth watching. No shorts and no clips. Filter by person to find what somebody appears in.

  • query Free text, matched against title, channel and the reason it is listed.
  • topic Research, People, Foundations, Safety, Silicon, Models, Policy, Agentic, Benchmarks.
  • person A person slug, e.g. andrej-karpathy.
  • limit

get_trending

Current trending items from one of four sources. Each ranks on a different metric and the numbers are not comparable across sources: Hugging Face uses its own trending score, GitHub uses stars, Reddit and Hacker News use points. The reply states which.

  • source required
  • limit

list_recent_changes

The index's own changelog: which entries moved, when, and why. Useful for answering what has changed since a given date.

  • limit

Resources

Static context an agent can pull in once rather than asking for repeatedly.

pyyan://abouttext/markdown

How the index is built, what the sourcing levels mean, and what the numbers can and cannot be used for.

pyyan://categoriesapplication/json

The categories entities are filed under, with what each ranks on.

pyyan://about is worth reading first. It explains what the sourcing levels mean, what "Not published" means as a specification value, and what this data cannot be used for.

Prompts

brief_me_on_ai

A briefing on what has happened in AI recently, built from sourced stories rather than recollection.

  • since ISO date to start from, e.g. 2026-09-01.

compare_models

Compare two models on price, context, licence and what each is best for, with every figure attributed.

  • a required First model name or slug.
  • b required Second model name or slug.

How to use what comes back

Every figure carries the date it was checked. A specification is an object with a value, the source URL it came from, an asOf date and a confidence of high, medium or low. This is not a live feed, and quoting a price without its date is how a stale number becomes somebody's decision.

News items carry a sourcing level. primary means the link goes to the company, court, paper or filing itself. corroborated means two independent outlets and no primary source. Where sources disagreed, the story says so rather than picking one.

Ranks are editorial. They are a judgement, not a computed score, and each category record says what its ranking is based on.

What it will not do

There are no write tools. Nothing here creates, updates or deletes anything, every tool is annotated readOnlyHint, and the server holds no state between calls. Private data is not reachable: subscribers and short links have no tool, no resource and no route.

It is also not a search engine for the open web. It answers from this index and nothing else, which is the point: an agent gets an answer it can attribute to a source and a date rather than to a model's memory.

Also available as plain REST

The same data is at the REST API, with an OpenAPI 3.1 document at /api/v1/openapi.json, for anything that is not an MCP client. Both are free, both are read only, and neither needs a key.

Server name pyyan, version 1.0.0. If something is wrong, the source is in the response and corrections are welcome.