docs
browser-memory docs
An open catalog of pre-built web tools for AI agents: one tool per action per site, each carrying the exact requests and selectors that action needs. Connect the MCP server once and your agent has all of them, so it stops working out every page.
What is browser-memory?
Without a tool, a browsing agent pays the same tax on every run: take a screenshot, snapshot the DOM, guess selectors, retry, and burn tokens re-deriving a page someone else already figured out.
browser-memory removes that tax. Each tool is one action on one site, resolved once and then replayed: the exact request or the exact selectors, with a name, its parameters and what it returns. The agent calls it and gets structured JSON instead of a page to read. When a site ships new markup and a tool breaks, it is fixed once in the catalog, for everyone.
The system at a glance
The MCP server runs on your machine and never holds a session: the call it replays runs in your browser, against the cookies already there.
The pieces
| Piece | What it is | Where |
|---|---|---|
| MCP server | Runs on your machine and exposes the catalog to your agent: discover lists a site’s tools, run executes one and returns its data. | The local MCP server |
| Tool catalog | Hosted, read-only registry of curated tools, one per action per site. Nothing to install. | api.browser-memory.com |
| A tool | The unit of memory: one action on one site, with a name, its parameters, what it returns and whether it writes anything. | what it is · how it runs |
llms.txt | Plain-text index of the whole catalog. Zero auth, zero install — an agent can just fetch it. | /llms.txt |
Which way in?
| If you… | Use |
|---|---|
| want every pre-built action for a site available to your agent at once | MCP server |
| want to know which actions a site already has | The catalog |
| just want to read what the catalog knows, no install | llms.txt |
| need a site or action the catalog does not cover yet | request it |