npm package for AI SDK 7
snaprender-ai-sdk adds SnapRender, the screenshot API for AI agents, to the Vercel AI SDK. takeScreenshot lets the model see a web page on desktop, a phone or a tablet, and extractContent lets it read a web page as markdown. Pass snaprenderTools() to generateText, streamText or an agent, and the screenshot reaches the model as an image it can look at.
200 free captures a monthNo cardVersion 0.2.0, MIT
import { generateText, isStepCount } from 'ai';
import { snaprenderTools } from 'snaprender-ai-sdk';
const { text } = await generateText({
model: 'anthropic/claude-sonnet-5.5',
prompt: 'Take a screenshot of https://vercel.com on an iPhone and tell me what the hero section says.',
tools: snaprenderTools(),
stopWhen: isStepCount(4),
});
console.log(text);import { takeScreenshot, extractContent } from 'snaprender-ai-sdk';
const tools = {
takeScreenshot: takeScreenshot(),
extractContent: extractContent({ timeoutMs: 30_000 }),
};npm install snaprender-ai-sdk ai zod
export SNAPRENDER_API_KEY=sk_live_...You need a SnapRender API key (free), Node.js 20.3 or newer and AI SDK 7. No key yet? Leave it unset: the first call starts a free trial key (100 watermarked captures) and each result includes a link to claim it. See agent trial keys.
Sign up with Google, GitHub or email. The key is on your dashboard right away and starts with sk_live_. The free plan has 200 captures a month (25 until you verify your email; Google and GitHub sign-ins start at 200), no card.
npm install snaprender-ai-sdk ai zod. ai and zod are peer dependencies: AI SDK 7 (ai@^7) and zod 3.25.76+ or 4.1.8+.
export SNAPRENDER_API_KEY=sk_live_.... The tools read SNAPRENDER_API_KEY, or pass { apiKey } or { getApiKey } to any factory.
tools: snaprenderTools() in generateText, streamText or a ToolLoopAgent, with a stopWhen that leaves room for the tool steps.
snaprenderTools() returns both. Import takeScreenshot or extractContent on its own when you need only one; each factory takes the same options.
See a web page
takeScreenshotCaptures a public URL or raw HTML and sends the image to the model. Desktop by default, or a phone, tablet or laptop with device; one screen, or the whole page with full_page.
iphone_14, iphone_15_pro, pixel_7, ipad_pro, macbook_pro.Read a web page as markdown
extractContentReads a page after its scripts run and returns the content as text: the full page as markdown, the main article, plain text, HTML, the links or the metadata.
selector reads one part, such as main.takeScreenshot inputsThe model fills these in. Pass url or html.
| Input | Type | Notes |
|---|---|---|
url | Typestring | Public http or https URL. Pass url or html, not both. |
html | Typestring | Raw HTML to render instead of a URL, up to 2 MB. |
format | Typejpeg | png | webp | Default jpeg at quality 80, the smallest file for the model. |
full_page | Typeboolean | The whole scrollable page instead of the first screen. |
lazy_load | Typeboolean | With full_page, scroll first so lazy images further down load. |
device | Typeiphone_14 | iphone_15_pro | pixel_7 | ipad_pro | macbook_pro | A phone, tablet or laptop view. Omit for desktop. |
width, height | Typenumber | Viewport in CSS px: width 320 to 3,840, height 200 to 10,000. Default 1280 x 800. |
dark_mode | Typeboolean | Asks the page for its dark theme (prefers-color-scheme: dark). |
delay | Typenumber | Extra wait after load, 0 to 10,000 ms. |
quality | Typenumber | JPEG or WebP quality, 1 to 100. |
hide_selectors | Typestring | Comma-separated CSS selectors to hide, such as a chat widget. |
click_selector | Typestring | A CSS selector to click before the capture, such as a tab. |
block_ads, block_cookie_banners | Typeboolean | Both on by default. |
cache | Typeboolean | Accept a copy from the last 24 hours, which uses no capture. Default false: render now. |
extractContent inputs| Input | Type | Notes |
|---|---|---|
url | Typestring | Public http or https URL. |
type | Typemarkdown | article | text | html | links | metadata | Default markdown. article keeps the main text only; links lists the links; metadata returns the title, description and Open Graph tags. |
selector | Typestring | Read only part of the page, such as main or #pricing. |
max_length | Typenumber | Most characters returned, default 100,000. |
delay | Typenumber | Extra wait after load, 0 to 10,000 ms. |
block_ads, block_cookie_banners, cache | Typeboolean | As for takeScreenshot. |
The screenshot goes back as an image the model can look at, next to a one-line summary. Pages come back as text.
takeScreenshot returns the image through the tool's toModelOutput: a text summary plus an image file part. The tool's own result, in steps[].toolResults and in UI messages, is a JSON object with source, mediaType, bytes, width, height, cache, remainingCredits and image (base64).
Image size. Captures render at 2x device scale, so a 1280 px wide viewport gives a 2560 px wide image. When an image is over maxImageBytes (5 MB) or maxImageDimension (8,000 px), the tool leaves it out and tells the model why, so it can retry with full_page: false or a smaller viewport.
extractContent returns the page content as text, followed by a short source line with the type and the word count.
Errors. A failed request throws SnapRenderError with status, code and the API's message, such as SnapRender 422 TARGET_UNREACHABLE. The AI SDK passes it to the model as a tool error, so the agent can explain or try another URL. Failed renders do not use a capture.
{
type: 'content',
value: [
{ type: 'text', text: 'Screenshot of https://example.com (2560x1600 px, 161337 bytes, cache MISS, 172 captures left this period).' },
{ type: 'file', mediaType: 'image/jpeg', data: { type: 'data', data: '<base64>' } },
],
}Works with generateText, streamText, ToolLoopAgent and any provider that accepts images in tool results.
takeScreenshot(), extractContent() and snaprenderTools() take the same options object. You set these in code; the model never sees them.
| Option | Default | Notes |
|---|---|---|
apiKey | Defaultprocess.env.SNAPRENDER_API_KEY | Your SnapRender API key. |
trial | Defaulttrue | With no key, use a free agent trial key (100 watermarked captures over 7 days) and add a claim link to results. false throws MISSING_API_KEY instead. |
trialAgent | Default'snaprender-ai-sdk' | Name shown on the claim page. |
getApiKey | Defaultnone | An async function that returns a key at call time, for secret stores. |
baseUrl | Defaulthttps://app.snap-render.com | The API address. |
timeoutMs | Default90000 | Per request. Full-page captures of long pages can take a while. |
maxRetries | Default1 | Retries after a per-minute rate limit, waiting as long as Retry-After says, up to 60 s. Monthly quota errors are never retried. |
maxImageBytes | Default5000000 | takeScreenshot only. Larger images are left out and the model is told why. |
maxImageDimension | Default8000 | takeScreenshot only. Largest width or height in px sent to the model. |
fetch | Defaultglobal fetch | A custom fetch, for proxies and tests. |
Rate limits: a per-minute limit (HTTP 429) is retried maxRetries times after the Retry-After wait. A used-up month comes back to the model as a plain-words tool error and is never retried. The tools cannot reach localhost, private networks or pages behind a login; for your own site behind a firewall, see allow SnapRender.
One environment variable to set
Sign up free, copy the key from your dashboard into SNAPRENDER_API_KEY and pass snaprenderTools() to your model.
200 captures a month free, no card. Every plan works with the package.
The same two tools for LlamaIndex, and ready-made packages for LangChain and CrewAI. Each one uses your SnapRender API key; the LlamaIndex tools also start a free trial key when none is set.
LlamaIndex: llama-index-tools-snaprender. SnapRenderToolSpec gives an agent take_screenshot, which returns a text summary plus an ImageBlock the model can look at, and extract_content, which returns the page as markdown, article text, plain text, links or metadata. Both have async versions that async agents use on their own.
The inputs match the AI SDK tools (device, full_page, lazy_load, selector, max_length and the rest). The constructor takes api_key (default SNAPRENDER_API_KEY), timeout in seconds, max_retries, max_image_bytes, max_image_dimension and get_api_key.
Use any LlamaIndex LLM integration that accepts images in tool results. Call the tools directly too:
spec = SnapRenderToolSpec()
blocks = spec.take_screenshot(url="https://example.com", full_page=True) # [TextBlock, ImageBlock]
markdown = spec.extract_content(url="https://example.com", type="article")pip install llama-index-tools-snaprender
export SNAPRENDER_API_KEY=sk_live_...import asyncio
from llama_index.core.agent.workflow import FunctionAgent
from llama_index.llms.anthropic import Anthropic
from llama_index.tools.snaprender import SnapRenderToolSpec
agent = FunctionAgent(
tools=SnapRenderToolSpec().to_tool_list(),
llm=Anthropic(model="claude-sonnet-5-5"),
system_prompt="You can look at web pages with take_screenshot and read them with extract_content.",
)
async def main() -> None:
response = await agent.run(
"Take a screenshot of https://www.llamaindex.ai on an iPhone (device iphone_14) "
"and tell me what the first screen shows."
)
print(response)
asyncio.run(main())Screenshot, cache check and usage tools for Python and JavaScript agents, LangGraph included. pip install langchain-snaprender or npm install langchain-snaprender.
A tool package that lets a crew capture pages and check usage. pip install crewai-snaprender.
Batches, signed links, webhooks and usage too. The AI SDK, the OpenAI Agents SDK and Mastra connect to the hosted MCP server with your key as a Bearer token.
The packages are free. Each screenshot or extraction uses one capture from your SnapRender plan; failed renders and cached copies use none. Every screenshot summary tells the model how many captures are left, and when the month runs out the tool error says so in plain words.
| Plan | Price a month | Captures a month |
|---|---|---|
| Free | $0 | 200 |
| Starter | $9 | 2,000 |
| Growth | $29 | 10,000 |
| Business | $79 | 50,000 |
| Scale | $199 | 200,000 |
The free plan has 200 captures a month (25 until you verify your email; Google and GitHub sign-ins start at 200). Every plan also includes scheduled screenshots and change monitoring. Yearly billing costs ten months of the monthly price. All plans.
An npm package that adds SnapRender, the screenshot API for AI agents, to the Vercel AI SDK. It has two tools: takeScreenshot lets the model see a web page as an image, on desktop, a phone or a tablet, and extractContent lets it read a web page as markdown, article text, links or metadata. snaprenderTools() returns both, ready for generateText, streamText or an agent.
AI SDK 7 (ai@^7), with zod 3.25.76 or later, or 4.1.8 or later, and Node.js 20.3 or newer. Install all three with npm install snaprender-ai-sdk ai zod.
Any provider that accepts images in tool results, such as Anthropic, OpenAI and Google models. The screenshot reaches the model as an image file part next to a short text summary. extractContent returns plain text, so it works with any model that can call tools.
The package is free and MIT licensed. Captures come from your SnapRender plan: Free $0 for 200, Starter $9 for 2,000, Growth $29 for 10,000, Business $79 for 50,000, Scale $199 for 200,000 captures a month. The free plan has 200 captures a month (25 until you verify your email; Google and GitHub sign-ins start at 200). Each screenshot or extraction uses one capture; failed renders and cached copies (cache: true) use none.
The capture was over maxImageBytes (5 MB by default) or maxImageDimension (8,000 px). Captures render at 2x device scale, so a tall full-page capture can pass the limit that vision models accept. The tool leaves the image out and tells the model why, so it can try again with full_page: false or a smaller viewport. Raise the limits if your model accepts larger images.
A per-minute rate limit is retried once by default (maxRetries), after waiting as long as the Retry-After header says, up to 60 seconds. Monthly quota errors are never retried. Failed requests throw SnapRenderError with the status, code and the API's message, which the AI SDK passes to the model as a tool error, so the agent can tell you.
The package gives you two typed tools and sends screenshots to the model as images, with no MCP client to run. The hosted MCP server gives all 11 SnapRender tools, including batches, signed links, webhooks and usage, through @ai-sdk/mcp: see the MCP setup for the AI SDK.
No. The tools reach public pages only, not localhost, private networks or signed-in sessions. If a capture of your own site shows a security check, one firewall rule lets SnapRender through: see allow SnapRender on your site.
npm install snaprender-ai-sdk, set a free API key and pass snaprenderTools(). 200 captures a month, no card.