# agents can read the web. orsel lets them do things on it.

Agents read pages with scrapers and act through screenshots. It is slow, brittle and unsafe, and nothing says what an agent is allowed to do. This replaces that with typed actions, a permission model and a record.

## the operations

The same seven operations work over REST, the TypeScript and Python SDKs, the CLI and the MCP server.

- **find**: pages and actions for an intent, ranked together, with the reason for each result.
- **read**: a page as clean markdown. Everything returned is page-derived and untrusted.
- **act**: run an action and get a typed result. It succeeds only when its postconditions pass.
- **watch**: monitor a search, a page or an action result for changes.
- **explain**: why this action, which recipe version, which evidence.
- **approve**: a person decides a waiting write. Single use, expiring, bound to the exact payload.
- **revoke**: pull a recipe or a delegation.

## how a run goes

An action is described by an Action Manifest: typed inputs and outputs, a risk class (read, write, sensitive, irreversible, financial) and machine-checkable postconditions. A recipe says how to run it, fastest tier first: official API or WebMCP, replayed first-party requests, a DOM script, then an agent as the last resort. Recipes are signed data, never trusted code.

Writes are denied by default. A policy can allow them, or require a person to approve the exact payload first. Irreversible and financial actions always need a person. Every find, run, decision and approval is appended to a hash-chained ledger, and `verify` recomputes the chain.

## your own browser

Logged-in actions run in your own browser, through the extension or the CLI, so your sessions stay on your machine. Hosted runs cover public pages only. Captchas and two-factor prompts are handed to you; they are never solved or bypassed.

## status

Early. The API, the SDKs, the CLI and the MCP server exist and are tested. Nothing is published to npm, PyPI or a package registry yet.
