Docs for AI agents

These docs are written for people and for agents. Everything a reader sees on a page is also available as plain markdown, and the whole corpus is indexed in a single file. Nothing here needs an API key or a browser.

Read a page as markdown

Append .md to any docs URL:

The response is text/markdown. It starts with the page title, its one-line description, a _Source:_ line pointing back at the HTML page, and the date a human last checked the page for accuracy. Links and images are absolute, so the markdown stands on its own once it is out of context.

Every page has one. On the page itself, Copy page in the toolbar puts the same markdown on your clipboard, and View as Markdown opens it.

Or ask for markdown with Accept

If your client sends an Accept header, you do not need the .md suffix. A request for a docs page that prefers text/markdown gets the markdown twin from the page URL itself:

text/markdown has to be preferred at least as much as text/html, so a browser, or curl with its default Accept: */*, still gets HTML. Negotiated responses are sent with Cache-Control: private, no-store and Vary: Accept; fetch the .md URL when you want a response you can cache.

Every HTML page also advertises its twin in the response headers, so a crawler that fetched the HTML can find the markdown without guessing:

Index the whole corpus

Two generated files cover the full set:

Both follow the llmstxt.org convention. Start with llms.txt to decide what you need, then fetch the individual .md files — llms-full.txt is there for the cases where one request is easier than many.

Search the docs without an LLM

The search box on these pages runs over a static JSON index you can fetch yourself:

It lists every visible page (path, title, navTitle, description, keywords, section and group) and every ## and ### heading (path, anchor, text). An API reference endpoint page is a single entry — its method and path are in keywords — so its headings are not listed. Build https://taostats.io/docs/{path}#{anchor} from an entry to link straight to an answer. The index changes only when the docs are deployed, so it is safe to cache for a few minutes.

Every page lives here now, the API reference included, so nothing in these docs links to docs.taostats.io.

Heading anchors are stable

Every ## and ### has an id generated from its text, so you can cite a specific answer rather than a whole page:

The same anchors appear in the .md twin, which makes {page path}#{anchor} a usable chunk identifier for retrieval.

Live data: the MCP server

Docs explain how Bittensor works. They cannot tell you a wallet's balance, this epoch's emission or which subnets are registering — those change by the block. For that, connect to the Taostats MCP server:

It exposes the Taostats API as Model Context Protocol tools, so an agent can look up accounts, subnets, extrinsics, prices and staking state without you writing a client. You need a Taostats API key; see Model Context Protocol (MCP) for the setup, and Getting started with the Taostats API to create the key.

Building a client instead

If you are writing code rather than driving a chat model, the TypeScript SDK wraps the same API with types, and adds direct RPC access to the chain for transactions. The REST API base URL is https://api.taostats.io/api; every endpoint is documented in the API reference.

API key hygiene

A Taostats API key identifies your account and counts against your plan's limits. Treat it the way you would treat any other credential:

• Keep it in an environment variable (TAOSTATS_API_KEY), never in a committed file, a prompt, or a rule file that ships with a repository.
• Do not paste it into a chat window. An agent that needs it should read it from the environment or from your MCP client's own secret store.
• Give an agent its own key, so you can rotate or revoke that key without breaking everything else.
• Rotate keys from Taostats Pro if one is exposed.

Rule-file snippets

Dropping a few lines into your repository's agent instructions saves an agent from crawling HTML. The wording is the same in each case — only the file name changes.

CLAUDE.md or AGENTS.md (Claude Code, Codex, and most agent CLIs):

.cursor/rules/taostats.mdc — the same content with Cursor's frontmatter:

.github/copilot-instructions.md — Copilot has no import mechanism, so paste the AGENTS.md block above into it verbatim.

Prompt snippet

For a one-off conversation, this is enough to orient a model:

What not to do

• Do not scrape the HTML. The .md twin is the same content without the navigation, and it will not break when the layout changes.
• Do not cache a page forever. Each page carries the date it was last reviewed; dTao, hyperparameters and API surfaces all move.
• Do not treat an example value as current. Sample balances and block numbers in these docs are illustrations, dated where it matters.