What is Manifest?
Manifest is an API that transforms any webpage into a structured JSON map of interactive elements an AI agent can click, fill, and submit. Instead of relying on fragile CSS selectors or screenshots, it reads the page's accessibility tree and cross-references it with the DOM to produce a clean action manifest. Each action comes with resolved locators and a requires field that encodes dependencies between elements — for example, "select a plan before this button becomes clickable." This gives agents a reliable, machine-readable blueprint of what they can actually do on a page.
Who it's for
- AI agent developers who need a reliable way to programmatically interact with any webpage without hand-writing or maintaining brittle selectors.
- Automation engineers building browser-based workflows that must survive UI redesigns and work across different sites without constant maintenance.
- Product teams shipping browser agents who want their agents to reason about page actions rather than guess from screenshots or raw HTML.
Key features
Action dependency mapping
Every interactive element in the manifest includes a requires field that encodes dependencies between actions. This tells an agent which fields or steps must be completed before a button becomes clickable or a form can be submitted — context that accessibility snapshots alone don't provide.
Resolved locators
Each action comes with resolved CSS and role locators, eliminating the need for developers to write or maintain selectors. The API handles the fragile work of identifying elements, so agents can reliably find and interact with them even after page redesigns.
Structured JSON output
The API returns a clean, predictable JSON manifest containing every button, form, and input on the page, along with their types, required fields, placeholders, and disabled states. This structured format lets agents reason about available actions and make informed decisions about what to do next.
Multi-platform support
Manifest ships with a Python SDK, LangChain support, and an MCP server, making it easy to integrate into existing agent frameworks and workflows across different programming environments.
What stands out
"Manifest is the layer that tells an agent what to click, fill, and submit — and what's mandatory before it can move on."
While browser tools give agents access to a page and content tools extract text, Manifest fills the missing layer: actionable intelligence. It doesn't just tell an agent what's on the page — it tells it what it's allowed to do there, what each action requires, and what dependencies must be satisfied first. This turns a webpage from a visual or textual document into a structured action space an agent can navigate autonomously.
Worth checking out if…
You're building browser agents that need to interact with real webpages reliably, without fragile selectors or screenshot-based guessing. Manifest is especially valuable if your agents must handle forms, multi-step workflows, or pages that change frequently — and you want them to reason about dependencies and required fields the way a human would.
