Retriever AI: run browser tasks from Chrome, cloud workflows or an API
Retriever AI is a browser agent for gathering information and completing tasks across websites. Its official site describes work on both the open web and sites where you are signed in, with options ranging from a Chrome extension to cloud runs and developer integrations.
Start from the page you already use
The browser extension is aimed at work that begins in an existing session. Examples include turning listings into a dataset and collecting information from multiple pages. The site also shows outreach and social publishing examples, which involve actions as well as reading.
A useful first task is a small, read-only extraction with a clear output shape. Ask for a few records, retain their source links and compare the fields with the original pages. That gives you evidence about the workflow before expanding it to messages, account changes or scheduled posting.
Choose how the task runs
Retriever offers five entry points: a browser extension, cloud execution, an API, MCP access and a CLI or SDK. The site describes scheduled cloud work and an API returning structured JSON. MCP is presented as a way to give assistants browser access from their existing conversation.
These are different ways to invoke the agent, not proof that every site or task behaves identically in each environment. In particular, check which account is signed in, where the browser runs and what information is returned when a task stops partway through.
Keep output checks close to the task
The homepage advertises a free, ad-supported option. It also presents customer counts, benchmark rankings and average task costs; those are publisher claims rather than results from our testing. The captured page is not a complete pricing or permissions specification.
For recurring work, define the expected output and a failure condition before scheduling it. A dataset should include enough source information to spot missing rows or stale values. For actions that send messages or change records, use an approval point and inspect a sample result. Retriever's main appeal is the range of browser-work entry points, but a dependable routine still needs a clear finish condition.