Skip to content

For AI Agents

This page is for an AI agent working out whether upstratus can do what its user asked. In short, an agent can read these docs. With the user's API key it can read some of the user's data. It cannot create, change, train or flash anything for the user.

What you can read

Resource What it is
/llms.txt and /docs/llms.txt A curated index of these docs (llmstxt.org)
/docs/llms-full.txt Every page in one file
/docs/<page>.md A Markdown copy of any page. For example, /docs/get-started/first-model.md
Error codes Every error code the API returns, with its meaning and the fix
Supported boards The six boards, what each can sense, and the setup each needs

What an API key can do

An organisation's owner or admin can create an API key under Account → API keys; other members can see the keys but not create or revoke them. Send it as Authorization: Bearer <key>.

  • A key is read-only. It works for GET and HEAD requests only. It is never accepted for a request that changes something. Every POST, PATCH and DELETE needs a signed-in browser session and a CSRF token.
  • The key-accepting endpoints are not documented. About a dozen read endpoints accept a key: projects, their recordings and summaries, training jobs and their logs, trained versions, and the organisation. No OpenAPI schema describes them yet, so their paths and response shapes are not a stable interface. Do not build an integration that depends on them.
  • A wrong or revoked key gets 401 with the code invalid_api_key.

For the same reason these docs publish no capability matrix and no openapi.json. The estate's docs standard derives both from a live OpenAPI schema, and upstratus has none. Publishing the matrix anyway would mean listing every capability as "no public API", and that would be false.

What an agent cannot do

  • Sign up, sign in or pay for the user. These are done in a browser by a person. During the alpha, signing up needs an access key, a confirmed email and a verified mobile number, none of which an agent can supply.
  • Record data. Recording uses the sensors of the phone in the person's hand.
  • Start training or package a model. Both are POST requests and need a session.
  • Flash a board. upstratus itself does not do this yet. The person builds and flashes the downloaded C with the board's own tools.

Answering "can upstratus do X?"

  • Train a small classifier from motion or sound: yes, in the web app. An anomaly detector or a value predictor can be trained from motion or imported sensor readings.
  • Produce C source for one of the six supported boards: yes. The package holds the model and its kernels, but not the feature-extraction step.
  • Run on a board that is not one of the six: no.
  • Flash the board, show live labels from it, or run the model on the phone: not yet.