Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@SlimSnap MCP look at my last capture"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Your recent screen captures, as structured data your AI can actually read.
Take a capture in SlimSnap, mark what you mean, then in Claude or Cursor say "look at my last capture". No pasting, no file paths, no clipboard.
Not a screenshot. A JSON document: every piece of text on screen with its position and colour, plus the arrows, boxes and callouts you drew, and which element each one points at.
That matters for three reasons.
A screenshot pasted into an agent gets downscaled before the model reads it, and nobody tells you. Anthropic's own docs say to pre-resize and crop your images because resizing "might make text less legible". A tall page loses its text entirely.
Images returned over MCP are worse still. In Claude Code they are passed through as text rather than native image blocks, costing roughly 15-25k tokens against about 1.6k for the same image pasted directly, and the model cannot read them at all (claude-code#31208, closed as not planned). A big enough result blows the per-result token ceiling outright (#9152).
And an oversized screenshot can poison a session permanently. It stays in context, gets re-sent every turn, and there is no way to remove it (#55040).
This connector sidesteps all three by never sending the picture. An ordinary screen costs roughly a third less than the same screen as an image, and on a tall page the comparison stops being about cost at all: shrinking an 18,410 pixel page to fit the model's 1568 pixel limit leaves body text 1.4 pixels tall, so the cheap image is cheap because it is unreadable. Either way the agent gets exact coordinates rather than a guess about which button you meant.
When you genuinely need pixels (visual style, spacing, colour balance) there is a separate tool for that, and the image is already downscaled to a size the model can read.
Related MCP server: Desktop MCP
Claude Code
claude mcp add slimsnap -- npx -y @slimsnap/mcpThen check it: /mcp should list slimsnap with 4 tools.
Claude Desktop
Download slimsnap-mcp.mcpb from Releases and open it. Claude shows an install dialog. The connector then appears under Settings.
Cursor
Or add it by hand to .cursor/mcp.json in your project, or ~/.cursor/mcp.json for every project:
{
"mcpServers": {
"slimsnap": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@slimsnap/mcp"]
}
}
}Project config wins over global when both exist. Then ask Cursor about your last capture the same way you would ask Claude.
As a Claude Code plugin (same connector, installed and updated as a plugin)
/plugin marketplace add bickov/slimsnap-mcp
/plugin install slimsnap@slimsnapYou need SlimSnap 0.5.0 or newer, because that is the version that started keeping recent captures on disk.
The first time Claude Code uses each tool it asks for permission, and the approval is scoped to the folder you were in. Four tools means four prompts, and they come back in the next project. Nothing is wrong; that is how Claude Code gates MCP tools.
To approve them once, everywhere, add this to ~/.claude/settings.json:
{
"permissions": {
"allow": [
"mcp__slimsnap__get_latest_capture",
"mcp__slimsnap__list_captures",
"mcp__slimsnap__get_capture",
"mcp__slimsnap__get_capture_image"
]
}
}If the file already exists, merge the allow entries into it rather than replacing it. Restart Claude Code afterwards.
Tool | What it does |
| The newest capture as structured JSON. This is the one that gets used most. |
| Recent captures newest first, with a text preview. Use it to find an older one, or to work through several marked screens at once. |
| One capture by id. |
| One frame as a downscaled image, for when the JSON cannot answer the question. |
Four tools on purpose. Tool descriptions are charged against your context window on every turn, so the surface stays small.
Two sample captures live in samples/ — one single screen, one
three-frame scroll capture, both with an arrow resolved to the element it points
at. Copy either into ~/.slimsnap/recent/ and the connector will read it, so you
can try all four tools before installing SlimSnap.
mkdir -p ~/.slimsnap/recent
cp -R samples/single-frame ~/.slimsnap/recent/2026-07-30T09-00-00-000000-00-00Then ask your agent what did I mark in my last capture? — it should come back with the Save changes button and the note attached to it.
SlimSnap keeps your recent captures in ~/.slimsnap/recent/, newest 20 by default. Each capture is a folder with capture.json, a small meta.json index, and one PNG per frame capped at 1568px on the long edge.
The connector finds that folder by reading ~/.slimsnap/config.json, which SlimSnap rewrites whenever your settings change. Move your save folder and the connector follows, with nothing to reconfigure.
Change how many are kept, or turn the cache off entirely, in SlimSnap under Settings, Export, Recent captures. There is a Clear button in the same place.
The server reads files under your home directory and talks to nobody. There is no network code in it. It cannot capture your screen and it cannot ask SlimSnap to capture your screen. It only reads captures you already took and chose to keep.
Nothing is uploaded, there is no account, and no telemetry.
No personal data is collected, stored, transmitted or shared. The server has no network code and no analytics. Full policy: https://slimsnap.ai/privacy.
SlimSnap 0.5.0 or newer, macOS or Windows. Works with Claude Code, Claude Desktop and Cursor.
Codex is not supported yet. It caps a single tool result at about 10 KB, which a real capture goes past, and it drops image content when structured output is present (codex#7906, #10334). Both are open upstream.
The Claude Desktop route needs nothing else: Desktop ships its own Node and the .mcpb bundle uses it. The Claude Code route runs node from your PATH, so it needs Node 18 or newer installed. That catches people out on a fresh Windows machine, where Node usually isn't there.
MIT
Returns one specific capture as structured JSON, same shape as get_latest_capture. Get ids from list_captures.
Inputs: id (required).
Returns the actual pixels of one frame, already downscaled to a size the model can read. Useful for visual questions that the text and coordinates cannot settle, such as spacing, visual style, colour balance, or imagery with no text in it. Costs substantially more tokens than the structured form of the same capture.
Inputs: id, frame.
Returns the newest SlimSnap capture as structured JSON: every OCR'd text element with its normalised bounding box and colour, plus any annotations the user drew (arrows, boxes, highlights, callouts) and which element each one points at. Use this whenever the user refers to their screen, their last screenshot, 'this', 'here', or something they just marked. Returns no image. The text and coordinates answer most questions about a capture on their own; the pixels are available separately when a question is genuinely visual.
Inputs: no named parameters.
Lists the user's recent captures newest first, with a short text preview of each. This is the cheap summary view; get_capture returns a single capture in full. Useful for finding an earlier capture when the user does not mean the newest one, or for working through several marked screens.
Inputs: limit. Returns: count, captures
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 4 tools
Each tool has a clearly distinct role: list_captures gives an overview, get_latest_capture and get_capture return structured data for different selections, and get_capture_image provides the raw pixels. The descriptions reinforce these boundaries, so an agent should not confuse them.
All tool names follow a consistent verb_noun pattern: list_captures, get_capture, get_latest_capture, and get_capture_image. The variation in pluralization and the additional 'latest' qualifier are natural and predictable.
Four tools is a well-scoped set for a capture retrieval server. There are no redundant tools and each one contributes a distinct function, covering both listing and accessing individual captures.
The tool surface fully covers the core lifecycle of reading captures: listing available captures, retrieving the latest capture, retrieving a specific capture by ID, and fetching the raw image when needed. There are no obvious missing operations for the stated purpose.
Screenshots, PDFs and Markdown from any URL or HTML for AI agents, via the SnapForge API
SnapRender gives AI agents eyes on the web through a hosted MCP server. Send a public URL and get the page back as an image, a PDF or clean markdown, with no browser to run or maintain. Capture the viewport or the full page as PNG, JPEG, WebP or PDF, at desktop size or with a phone, tablet or laptop preset, in light or dark mode. Ads and cookie banners are removed by default, and you can hide elements or click one before the capture. extract_content reads the same page as markdown, text, article, links or metadata, so an agent can see a page and read it in one session. Made for monitoring and client work: batch up to 50 URLs in one job, check the cache for free before spending a capture, share signed links that render without an API key, and get webhooks when a batch finishes or a scheduled capture spots a change. Track competitors' pricing and landing pages, or send clients current screenshots of their sites. Connect with OAuth (Google, GitHub or email) or an API key. Works with Claude, ChatGPT, Cursor, VS Code and any MCP client. Free plan: 200 screenshots a month, no credit card. Paid plans from $9 a month for 2,000 up to $199 a month for 200,000. Home: https://snap-render.com Docs: https://snap-render.com/ai-agents
Screenshot, PDF and HTML-to-image rendering API so Claude and Cursor can see any web page.
Screenshot, PDF and HTML-to-image rendering API so Claude and Cursor can see any web page.
