Dependency and CI trust policy, local audit commands, and scorecard promotion
rules are documented in Supply-chain policy.
Plasmate compiles HTML into a Semantic Object Model (SOM), a structured
representation that LLMs can reason about directly. It runs JavaScript via V8
and implements a documented CDP subset for supported Puppeteer workflows. SOM
output removes presentation and runtime markup; output size and tokenization are
page-, configuration-, serialization-, and tokenizer-dependent.
Browser engine for agents. HTML in, structured Semantic Object Model out. Native MCP and AWP with a supported CDP subset. Apache-2.0.
Yıldızlar
38
Araçlar
—
Çalışma süresi
—
Plasmate
Lightpanda
Chrome
Primary output
Structured SOM JSON
Raw HTML / DOM
Raw HTML / DOM
JavaScript
Embedded V8 pipeline
Browser runtime
Browser runtime
Agent protocol
Native MCP and AWP
CDP
CDP
Benchmark policy
Retained, denominator-complete reports
Project-specific
Project-specific
License
Apache-2.0
AGPL-3.0
Chromium
Install
curl -fsSL https://plasmate.app/install.sh | sh
Or install the native engine with Cargo:
cargo install plasmate
The npm and PyPI packages named plasmate are client SDKs; they do not bundle
or install the native plasmate executable. Install the native engine first,
then add npm install plasmate or pip install plasmate to applications that
need the corresponding SDK.
Quick Start
Fetch a page and get structured output
plasmate fetch https://news.ycombinator.com
Returns SOM JSON: structured regions, interactive elements with stable IDs, and content. Measure output size on the pages and configuration used by your application.
AWP's foundational v0.1 core has seven wire methods: awp.hello,
session.create, session.close, page.navigate, page.observe, page.act,
and page.extract. The current native handler also exposes session-listing,
network-interception, plugin, cookie, and proxy-pool extensions. Click, type,
select, scroll, toggle, and clear are page.act actions, not separate wire
methods.
Plans are versioned and bounded; mutating steps require separate
--confirm-step <id> approvals, while JavaScript evaluation and cookie writes
also require category opt-ins. See the workflow contract
for schemas, secret references, containment, and execution examples.
Run as an MCP tool server (Model Context Protocol)
plasmate mcp
This exposes Plasmate over stdio as MCP tools:
fetch_page - get structured SOM from any URL
extract_text - get clean readable text
extract_links - get deduplicated links from a page
ard_discover - inspect bounded static ARD v0.9 draft catalog signals without invoking them
The endpoint is http://127.0.0.1:9272/mcp. Every request must include
Authorization: Bearer <token>. The server implements JSON response mode for
the stateful MCP 2025-11-25 transport and the stateless 2026-07-28 release
candidate. It deliberately returns HTTP 405 for GET rather than claiming an
SSE notification stream it does not implement. See
docs/mcp-streamable-http.md for lifecycle,
headers, and browser-Origin policy.
Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
Cursor — ~/.cursor/mcp.json
VS Code Copilot — .vscode/mcp.json (workspace) or user settings
Windsurf — ~/.codeium/windsurf/mcp_config.json
Once connected, the native server advertises its current tool surface through
MCP tools/list. It includes stateless fetch/extraction/discovery tools,
cache/session/trace inspection, screenshot and stateful page interaction,
cookie operations, and validation-only replay. Query tools/list instead of
depending on a hard-coded count; the authoritative registration is
src/mcp/server.rs.
Tip: use selector="main" (or main_content/main_contents/main-content/main-contents/main content/main contents/main_region/main_regions/main-region/main-regions/main region/main regions) to strip nav/footer, selector="interactive"
to return only actionable elements, selector="h1" through selector="h6"
(or heading1 through heading6) to isolate a heading level, or an action selector such as selector="click"
(the equivalent selector="action:click" also works), selector="type",
selector="clear", selector="select", selector="toggle",
selector="submit", or selector="reset" (equivalently
selector="action:reset", including action : reset and action = reset)
to build a compact
action menu before the LLM sees the content. The MCP tool-name aliases
selector="type_text" and selector="select_option" (including their
action:type_text and action:select_option forms) are also accepted for
typing and selection targets.
Use selector="label:<text>" or selector="label=<text>" for
case-insensitive accessible-label/text matching on regions and elements;
element matching also falls back to aria-label, title, placeholder, and
alt attributes when a compiled label is unavailable;
whitespace is normalized so labels split across lines still match.
Element selectors also accept common HTML tag aliases: a/link,
p/paragraph, and ul, ol, dl, or menu/list/lists (plus the explicit
unordered_list, ordered_list, and definition_list forms).
Structural aliases include fieldset/fieldsets/group/groups, ARIA group roles
figure/figures/radiogroup/radiogroups/progressbar/progressbars/
meter/meters/output/outputs, hr/separator,
details/summary/summaries, and iframe; img/image and ARIA widget names such
as menuitem/menu items, option, tab, treeitem/tree items, switch,
menuitemcheckbox, and menuitemradio are also accepted. Region selectors accept ARIA landmark
aliases such as navigation, nav region/nav-region/nav regions,
navigation region/navigation-region/navigation regions, search/search navigation/search navigations, banner, complementary, and
contentinfo; forms and dialogs are plural aliases for form and
dialog; content_region/content_regions (and their hyphenated or
space-separated forms) select content regions.
Table selectors include table/tables, data_table/data_tables, and
grid/grids/data_grid/data_grids/treegrid/treegrids (with
hyphenated and space-separated forms accepted too).
Use selector="media" to keep images and media-capable grouped elements
such as audio, video, and canvas while removing unrelated page content.
HTML text-input type names are also accepted as selectors, including
text, search_input/search-input, search_control/search-control,
search_field/search-field, search_box/search-box, query,
query_input/query-input, query_field/query-field,
query_control/query-control, querybox/query-box, email, number,
password, tel, url, date,
datetime/datetime-local (also date_time/date-time/date time and date_time_local/date-time-local/date time local), month, time, week, color, range,
hidden, and file; each targets the corresponding compiled text-input
element.
Use cache_status after repeated fetches to inspect local MCP SOM cache hits,
misses, selector entries, effective-HTML entries, and avoided HTML work.
Use session_status before long interactive runs to inspect active browser
session count, capacity, loaded URLs, raw/effective HTML sizes, SOM sizes,
node-map counts, structured data presence, disabled/readonly interactive
counts, age, and idle time. Stateful
open_page and navigate_to return cache_restored=true when they reuse a
content-hash-validated cache entry with both SOM and effective HTML.
Set trace=true on open_page to opt into bounded, memory-only action tracing,
then use trace_status, trace_export, and side-effect-free
replay_validate. Typed values and page bodies are never exported. See
docs/session-tracing.md for the privacy contract,
bounds, drift classes, and validation-only limitation.
Before a multi-page crawl, use crawl_policy or the CLI equivalent:
The versioned report distinguishes an unavailable robots file (4xx, access
permitted by RFC 9309) from an unreachable one (network/5xx, access denied).
It is advisory metadata, not authorization, and does not silently alter
ordinary fetch or fetch_page. See docs/CRAWL-POLICY.md.
Use inspect_page when an agent may need pixels. Its default auto mode
always returns a bounded compact SOM and only attaches a screenshot for named
signals such as near-empty structure, canvas-heavy content, or image-map/image
controls. never recommends without capture; always is an explicit capture
request. Page JavaScript is off by default; javascript=true runs untrusted
page code through the supervised JavaScript worker before the separately
isolated screenshot process.
Rendering uses a sandboxed/CSP-constrained document with JavaScript
disabled, no unsafe file-access or browser-sandbox switches, a dead network
proxy, and process-tree cleanup. Plasmate does not interpret the image with a vision model. See
docs/STRUCTURED-VISUAL-FALLBACK.md.
Static ARD discovery (draft)
Inspect the three static Agentic Resource Discovery signals supported by
Plasmate—the well-known catalog, an HTML rel="ai-catalog" link, and a
robots.txt Agentmap directive:
The versioned plasmate.ard.discovery.v1 report validates bounded catalog
envelopes and labels all publisher metadata, inline data, trust manifests,
attestations, and signatures as untrusted and unverified. It does not search a
registry, invoke entries, fetch entry endpoints, follow nested catalogs, or
verify identity. HTTPS and same-origin catalog discovery references are
mandatory, and the ARD path cannot use Plasmate's private-network development
override. Standalone reports and fully wrapped MCP tool results each have an
independently measured 512 KiB serialized bound. See
docs/ARD.md for limits and protocol scope.
Vercel AI SDK
Use Plasmate via the AI SDK's built-in MCP client (AI SDK v4+):
npm install ai @ai-sdk/openai
import { experimental_createMCPClient as createMCPClient, generateText } from 'ai'
import { Experimental_StdioMCPTransport as StdioMCPTransport } from 'ai/mcp-stdio'
import { openai } from '@ai-sdk/openai'
const mcp = await createMCPClient({
transport: new StdioMCPTransport({
command: 'plasmate',
args: ['mcp'],
}),
})
const { text } = await generateText({
model: openai('gpt-4o'),
tools: await mcp.tools(),
maxSteps: 5,
prompt: 'Summarize the top 3 stories on news.ycombinator.com',
})
await mcp.close()
This wires the full Plasmate MCP tool set directly into any Vercel AI SDK agent. See Vercel AI SDK MCP docs for details.
The checked-in MCP Registry
metadata declares ghcr.io/plasmate-labs/plasmate:v0.6.0 as the runnable,
version-pinned OCI package for the next release candidate. This image is
not available until the immutable v0.6.0 release workflow publishes it. Only
after it is anonymously pullable and the metadata is published can Registry
clients launch it over stdio with the mcp subcommand. The eventual direct
invocation is:
docker run --rm -i ghcr.io/plasmate-labs/plasmate:v0.6.0 mcp
The npm and PyPI packages named plasmate are client SDKs, not server
executables. For a host-native installation, install the Rust binary above
and configure the command as plasmate mcp.
What is SOM?
The DOM was built for rendering. SOM was built for reasoning.
SOM strips layout, styling, scripts, SVGs, and boilerplate. It keeps structure, content, and interactive elements with stable IDs that agents can reference in actions.
Output-size evidence
The retained v0.5.1 public-web snapshots in
website/docs/coverage.json and
website/docs/coverage-js.json recorded
median serialized-byte ratios of 9.98x over 83 successful non-JavaScript inputs
and 9.32x over 82 successful JavaScript inputs, respectively, from 98 attempted
URLs in each run. These are historical, observational byte ratios—not universal
token savings, cost savings, latency, or task-success claims. Blocked and failed
inputs remain in the denominator. See the benchmarking policy
and public claim registry before comparing or citing
results.
JavaScript Support
Plasmate embeds V8 and executes page JavaScript, including:
Inline and external <script> tags
fetch() and XMLHttpRequest with real HTTP requests
setTimeout / setInterval with timer draining
DOM mutations (createElement, appendChild, textContent, innerHTML, etc.)
DOMContentLoaded and load events
Promise resolution and microtask pumping
The JS pipeline runs during plasmate fetch and CDP page.goto(). The resulting DOM mutations are serialized back to HTML before SOM compilation, so JS-rendered content is captured.
CDP Compatibility
In the historical campfire-commerce fixture from
Lightpanda's Puppeteer benchmark,
Plasmate passed the assertions described in the
March 2026 exploratory report. That
single fixture is evidence for the listed workflow, not full Puppeteer,
Playwright, CDP, or web-platform compatibility. Supported CDP methods include:
Plasmate.getInteractiveElements returns a full-tree action menu, including
nested and shadow-root targets. It accepts optional role, action, label,
exact, enabledOnly, replay lookup (value plus by, or direct id,
cacheKey, htmlId, testId), and offset/limit params so Puppeteer/
Playwright clients can request, for example, one page of enabled click targets
before calling an LLM. Each target includes SOM role names, deterministic
cache_key, html_id, test_id, enabled, and blocked_reason when
available.
DOM.querySelector / DOM.querySelectorAll walk the SOM-backed node tree in
document order and understand stable replay selectors such as #html_id,
#som_id, [data-testid="..."], [data-test-id="..."], common attributes
such as [name="..."], [href="..."], [type="..."],
[aria-label="..."], [aria-labelledby="..."], [role="..."], SOM role
names, text, and labels. Matching for text and labels is case-insensitive.
CDP DOM nodes expose replay attributes such as data-plasmate-id,
data-som-role, HTML id, test id, ARIA label, href/name/type, and
disabled/readonly/required flags when available. Accessibility.getFullAXTree
is also SOM-backed and includes nested and shadow-root elements with backend
node ids plus disabled/readonly properties when available.
CDP is a compatibility layer. AWP is the native protocol, designed for agents rather than debuggers.
Architecture
HTML → Network (reqwest) → HTML Parser (html5ever)
→ JS Pipeline (V8: scripts, fetch, XHR, timers, DOM mutations)
→ DOM Serialization → SOM Compiler → JSON output
Network: reqwest with TLS, HTTP/2, redirects, compression; cookie jar supported, cookie APIs and proxy configuration are still limited
JS Runtime: V8 with DOM shim (80+ methods), blocking fetch bridge
SOM Compiler: semantic region detection, element ID generation, interactive element preservation, smart truncation, deduplication
Protocols: AWP (seven foundational core methods plus current extensions)
and a CDP compatibility layer for supported Puppeteer workflows
Requirements: Rust 1.88+ for both default and plugins builds. V8 is fetched
automatically by rusty_v8. CI checks the declared minimum against default and
all-feature builds so dependency updates cannot silently raise it.
Docker
Prebuilt multi-arch images (linux/amd64 and linux/arm64) are published to GHCR:
# Server mode (CDP or AWP)
docker run --rm -p 9222:9222 ghcr.io/plasmate-labs/plasmate:latest
# One-shot fetch
docker run --rm ghcr.io/plasmate-labs/plasmate:latest fetch https://example.com
MCP cache/session observability for repeated agent workflows
Parallel sessions at scale (500+ concurrent)
Ecosystem
Plasmate documents integrations across the AI and developer ecosystem. The
links below are an ecosystem directory, not a count of currently tested or
maintained compatibility contracts. Verify each integration's repository,
version support, and test status before adopting it:
Henüz araç listesi yok. Uzak sunucular ilk sağlık kontrolünden sonra bir liste alır.
SSS
Sorular
Plasmate nasıl kurulur?
Bu sayfadan Cursor, Claude Desktop, Claude Code veya VS Code için stdio yapılandırmasını kopyalayın. PlugMCP proxy'sini değil, yayıncının kurulumunu kullanır.
Plasmate için API anahtarı gerekiyor mu?
Bu kayıtta gerekli bir ortam değişkeni listelenmiyor.
Plasmate hangi aktarımı kullanıyor?
Plasmate yerel olarak stdio üzerinden çalışır.
Plasmate en son ne zaman kontrol edildi?
Son kontrol: 2026-10-28. Browser engine for agents. HTML in, structured Semantic Object Model out. Native MCP and AWP with a supported CDP subset. Apache-2.0.