llms.txt Content
# Plasmate
> Open-source browser engine for AI agents. HTML in, structured Semantic Object Model out.
Plasmate is a Rust headless-browser engine that compiles web pages into a
Semantic Object Model (SOM): JSON with typed regions, interactive elements,
stable references, actions, and content. It runs as a CLI, daemon, native MCP
server, AWP server, and CDP compatibility server.
## Release status
- Latest published Rust engine: v0.5.1.
- Next candidate: v0.6.0; it is not released or available as an OCI image yet.
- The v0.6.0 MCP Registry declaration must not be published until its newly
labeled GHCR image is built and anonymously pullable.
- Apache-2.0 licensed; source: https://github.com/plasmate-labs/plasmate.
## Protocol surfaces
- MCP: the native server advertises its current tools through `tools/list`.
Capabilities include stateless fetch/extraction/discovery, cache/session/trace
inspection, screenshot and stateful page interaction, cookies, and
validation-only replay. Do not rely on a hard-coded tool count.
- AWP: the foundational v0.1 core methods are `awp.hello`, `session.create`,
`session.close`, `page.navigate`, `page.observe`, `page.act`, and
`page.extract`. The current wire handler also includes session-listing,
network-interception, plugin, cookie, and proxy-pool extensions. Primitive
click/type/select/scroll/toggle/clear operations are `page.act` actions.
- CDP: a compatibility layer for supported Puppeteer/CDP workflows.
## Output-size evidence
Output reduction is page-, configuration-, corpus-, and serialization-dependent.
Retained v0.5.1 observational public-web snapshots recorded:
- non-JavaScript: median serialized-byte ratio 9.98x over 83 successful inputs
from 98 attempted URLs;
- JavaScript: median serialized-byte ratio 9.32x over 82 successful inputs from
98 attempted URLs.
Blocked and failed inputs remain in the denominator. These byte ratios are not
universal token savings, cost savings, latency, or task-success