SnapStack Server
Enables MCP-capable LLM clients to read browser screenshots captured by the SnapStack extension, stored locally and served over Streamable HTTP.
README
<p align="center"> <img src="assets/logo.png" alt="SnapStack" width="440"> </p>
<p align="center"> <a href="https://github.com/bgaze/snapstack-server/actions/workflows/ci.yml"><img src="https://github.com/bgaze/snapstack-server/actions/workflows/ci.yml/badge.svg" alt="CI"></a> <a href="LICENSE"><img src="https://img.shields.io/github/license/bgaze/snapstack-server?color=blue" alt="License: MIT"></a> <img src="https://img.shields.io/badge/node-%3E%3D18-brightgreen" alt="Node >= 18"> <a href="https://modelcontextprotocol.io"><img src="https://img.shields.io/badge/MCP-compatible-blueviolet" alt="MCP compatible"></a> <img src="https://img.shields.io/badge/100%25-local-success" alt="100% local"> <a href="https://www.npmjs.com/package/snapstack-server"><img src="https://img.shields.io/npm/v/snapstack-server?label=npm" alt="npm version"></a> <a href="https://www.npmjs.com/package/snapstack-server"><img src="https://img.shields.io/npm/dm/snapstack-server?label=downloads" alt="npm downloads"></a> </p>
<p align="center"> <img src="assets/demo.gif" alt="SnapStack demo — capture a browser tab, your AI reads the screenshots over MCP" width="900"> </p>
The SnapStack server is a single always-on Node process: it receives browser captures from the
extension, stacks them on disk, and serves them to any
MCP-capable LLM client over Streamable HTTP. It listens only on 127.0.0.1 — nothing ever leaves your machine.
New here? The full install + usage guide lives in the extension README: snapstack-extension. This page is the technical reference.
Architecture
One always-on process serves both the extension (capture) and your MCP client, decoupled by a folder on disk.
[MV3 extension] --POST /push (bytes)--> ┐
▼
[SnapStack server] 127.0.0.1:4123
├─ writes ─► ~/.snapstack/ (stack on disk)
└─ MCP /mcp (HTTP) ◄── MCP client
- Capture — the extension encodes the shot as WebP (PNG fallback), downscales it, and POSTs it here.
- Stack — one image file (
.webp/.png) plus a twin.json(url, title, timestamp, dimensions) per capture, namedNN <timestamp>: a stable two-digit number (assigned in capture order, restarts at01when the stack empties) plus a timestamp, under~/.snapstack/. - Retrieval —
get_screenshotsreturns a JSON manifest (number, absolute path, dimensions, metadata — no image bytes); the client reads only the files it needs, by path. Deletion is a separate, explicitclear_screenshotsstep. Retrieval never deletes.
Requirements
- Node.js ≥ 18 (tested on Node 20). No git needed at runtime.
- An MCP-capable LLM client speaking the HTTP (Streamable HTTP) or stdio transport.
- The snapstack-extension loaded in your browser.
Install & run
Run it once in the foreground:
npx -y snapstack-server@latest # → SnapStack server listening on http://127.0.0.1:4123
For start-at-login + crash-restart + self-update, install the auto-start unit (launchd on macOS, systemd --user on
Linux, a logon scheduled task on Windows):
npx -y snapstack-server@latest install # register auto-start; uninstall with `… uninstall`
The unit runs a best-effort npm install --prefix <appDir> snapstack-server@latest then launches the locally installed
copy — so the server self-updates on each (re)start, and still starts offline once installed. No git involved.
The full end-to-end walkthrough (idiomatic install paths, MCP client registration, the extension) is in the extension README.
MCP
SnapStack speaks two MCP transports over the same on-disk stack — pick whichever your client supports:
// HTTP (server already running) — register http://127.0.0.1:4123/mcp; copy deploy/mcp.json
{ "type": "http", "url": "http://127.0.0.1:4123/mcp" }
// stdio (the client spawns the process)
{ "command": "npx", "args": ["-y", "-p", "snapstack-server", "snapstack-mcp"] }
The HTTP /mcp endpoint is stateless (a fresh server + transport per request); the stdio front-end (snapstack-mcp)
is spawned on demand and reads the same ~/.snapstack stack. Capture intake (/push) always stays in the running
server, independent of either MCP front-end.
Exposed tools
| Tool | Description |
|---|---|
get_screenshots |
Lists pending captures as a JSON manifest (stable number, absolute path, dimensions, metadata) — no image bytes, no deletion. Pass numbers (e.g. [1,3]) to list only those. |
clear_screenshots |
Deletes captures. Pass numbers to delete specific ones; omit to clear the whole stack. Numbering restarts at 01 once empty. |
count_screenshots |
Number of pending captures, without retrieving them. |
get_screenshots and count_screenshots are read-only; only clear_screenshots is destructive. To run a
tool without a per-call confirmation, add its identifier to your client's allow-list (for Claude Code:
mcp__snapstack__<tool> in permissions.allow).
Token cost:
get_screenshotsreturns only the manifest, so it stays cheap whatever the stack size — the client then reads just the files it needs. WebP + downscaling keep those reads light.
Configuration
Environment variables (infrastructure)
| Variable | Default | Purpose |
|---|---|---|
SNAPSTACK_DIR |
~/.snapstack |
Stack folder. |
SNAPSTACK_PORT |
4123 |
Listening port (always on 127.0.0.1). |
Capture policy (shared across your browsers)
The encoding/capture settings are owned by the server and stored in ~/.snapstack/config.json, so a single
edit applies to every browser running the extension. They are edited from the extension's options page — not
an environment variable — and fetched by the extension before each capture.
| Key | Default | Meaning |
|---|---|---|
format |
webp |
Image format: webp, png or jpg. |
quality |
0.85 |
Lossy quality (0–1; the extension UI shows it as a percentage). |
maxWidth |
1568 |
Downscale captures wider than this to this width in px (0 = no resize). |
maxSlices |
50 |
Full-page capture: hard cap on stitched slices. |
Two endpoints back it: GET /config returns the effective policy; POST /config validates and replaces it (host- +
CORS-guarded like every capture route). The file is a non-image, so a stack clear never touches it; deleting it just
restores the defaults above.
Troubleshooting
- "Capture server not started" (in the extension): start the server (
npm start) or check the auto-start. Test:curl http://127.0.0.1:4123/health. - Port already in use (
EADDRINUSE): setSNAPSTACK_PORTto another value. - The client doesn't see the tools: the server must run before the MCP client starts; check the config
(
type: "http", correct URL). Direct test:curl http://127.0.0.1:4123/count. - Inspect the stack:
ls ~/.snapstack(image files + human-readable.json).
License
MIT — see LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.