Agent Discovery
skills.md and llms.txt: how Telebase is built for AI agents to discover
An AI agent that needs to check a phone number has no way to read a sales page or sit through onboarding. Telebase solves this with two machine-readable files: llms.txt at the site root, which tells an agent what Telebase does and where to look next, and skills.md, the full operating manual an agent reads once before making its first call. Together they let an agent go from "does this API exist" to a working, authenticated lookup without a human integrating anything first.
What each file is for
Modelled on the emerging llms.txt convention, this is a short, link-heavy Markdown file at telebase.io/llms.txt that states what Telebase is, what it does not do, current coverage and pricing, then points to every content page on the site grouped by topic. An agent (or a person) can read it in seconds and decide whether Telebase is relevant to the task at hand.
The full guide lives at telebase.io/skills-md, addressed directly to the agent reading it: how to get the user's API key, how to normalise a phone number to E.164, the exact request shape, how to interpret every field in the response, and how to handle each documented error code. It is written as instructions, not marketing copy: "ask once," "never fabricate a result," "each call costs a credit, do not repeat it."
Why a self-contained response matters more than a nice interface
A human integrating an API can tolerate ambiguity: read the docs twice, email support, guess and check. An agent acting inside a live conversation cannot. The response reference is written so every field has an explicit type and an explicit meaning for every possible value, including null, rather than leaving an agent to infer intent from a nice-looking dashboard it will never see.
This is also why the SIM swap field is documented the way it is. skills.md tells the agent explicitly to say "no SIM-swap data available" when the field returns UNKNOWN, and never to say "no SIM swap detected", because those two statements mean different things and an agent that blurs them gives the person it's helping a false sense of certainty. See giving AI agents phone intelligence tools for the fuller pattern of exposing this as a callable tool.
What this replaces for the engineer wiring it up
Without these files, an engineer adding Telebase to an agent has to write the tool description, the error handling logic and the "how do I explain this field to the user" guidance themselves, translating a human-facing docs page into agent-facing instructions by hand. With skills.md, that translation is already done and can be pointed to directly or summarised into a tool definition with minimal rewriting.
Request early access