About snipper-mcp
snipper-mcp is the MCP server bundled with SnipperApp 3, a native macOS code snippet manager. It lets AI assistants manage a local snippet library, with changes stored in the app's SQLite database.
Key features
- 01Searches and reads snippets
- 02Creates, updates, and deletes snippets
- 03Organizes folders, tags, and attachments
- 04Accesses public Hub snippets
- 05Uses the app's local SQLite database
Use cases
- 01Find a saved code example by topic
- 02Save a function with folder and tags
- 03Organize snippets and attachments
- 04Search or import a public Hub snippet
MCP support
The bundled stdio MCP server exposes tools for managing snippets, folders, tags, attachments, workspaces, and public Hub content. It reads and writes the same local SQLite database used by SnipperApp 3.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
We haven’t collected a package or remote endpoint for this server. Follow the instructions in the project README.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMEsnipper-mcp — the MCP server inside SnipperApp 3

SnipperApp 3 is a native macOS code snippet manager. It ships its own Model Context Protocol server, so AI assistants can search, read, create and organise the snippets in your library. There is nothing to install: the server is a stdio binary bundled in the app.
This repository documents the server's interface. The server itself is part of the app (Swift, built on the Swift MCP SDK).
Requirements
- SnipperApp 3 (Mac App Store, free 7-day trial): https://apps.apple.com/us/app/snipperapp-3-code-snippets/id6757330954?mt=12
- macOS 15 or later (Apple Silicon or Intel)
Install in Claude Desktop (one click)
Download snipperapp-1.0.2.mcpb from the latest release and open it; Claude Desktop installs it as an extension. The bundle contains only a small launcher (mcpb/server/index.js) that finds the server inside SnipperApp 3 and hands over stdio. Without the app it still answers initialize and tools/list (the manifest in mcpb/server/tools.json), and every tool call returns a clear not-installed error; a Dockerfile runs that mode for registry introspection checks.
Setup (manual)
Claude Code:
claude mcp add snipper -- "/Applications/SnipperApp 3.app/Contents/MacOS/snipper-mcp"
Claude Desktop, Cursor (.cursor/mcp.json), Windsurf (~/.codeium/windsurf/mcp_config.json):
{
"mcpServers": {
"snipper": {
"command": "/Applications/SnipperApp 3.app/Contents/MacOS/snipper-mcp"
}
}
}
Full guide: https://snipperapp.com/docs/mcp-integration
Tools (25)
| Area | Tools |
|---|---|
| Snippets | search_snippets, get_snippet, create_snippet, update_snippet, delete_snippet |
| Folders | list_folders, create_folder, update_folder, delete_folder |
| Tags | list_tags, create_tag, delete_tag, add_tag_to_snippet, remove_tag_from_snippet |
| Attachments | list_attachments, get_attachment, create_attachment, delete_attachment |
| Organisation | list_workspaces, list_storages, list_groups, list_languages |
| Hub (public snippets) | search_hub, get_hub_snippet, import_from_hub |
Every tool carries a title and the MCP annotations clients use to decide what needs confirmation: the twelve read tools are marked readOnlyHint, the five delete/untag tools destructiveHint, and the three Hub tools openWorldHint.
The server reads and writes the same local SQLite database as the app, so anything an assistant saves appears in SnipperApp immediately. Nothing is sent to a server; the Hub tools call the public Hub API only when you use them.
Example prompts
- "Search my snippets for the retry-with-backoff helper and paste it here."
- "Save this function as a snippet in the Work workspace, folder Utilities, tagged postgres."
- "List my Swift snippets tagged networking."
Privacy Policy
Full policy: https://snipperapp.com/privacy
What the server collects. Nothing. snipper-mcp has no analytics, no telemetry and no crash reporting of its own. The MCPB launcher in this repository makes no network requests at all — it either hands stdio to the bundled binary or answers tools/list from a static file.
How your data is used and stored. The server reads and writes the same local SQLite database as SnipperApp 3, inside the app's own container on your Mac. Snippets, folders, tags and attachments never leave your machine unless you have turned on a sync storage yourself: iCloud (Apple CloudKit, your private database — the developer cannot read it) or GitHub Gist (your own account, under your GitHub credentials).
Third-party sharing. None by the server. The three Hub tools (search_hub, get_hub_snippet, import_from_hub) call the public SnipperApp Hub API, and only when an assistant invokes them; they send the search terms or snippet ID and nothing from your library. No snippet content is uploaded. Nothing is sold or shared with advertisers or data brokers.
Data retention. Your snippets live on your Mac and stay until you delete them; deleting them in SnipperApp 3 (or via delete_snippet) removes them from the local database and from any sync storage you enabled. There is no server-side copy for the developer to retain. Hub requests are not tied to an account and are not retained as a per-user history.
Contact. [email protected] — for questions, or to ask what is held about you and have it deleted.
This repository
mcpb/— the MCPB extension:manifest.json,icon.png, and the launcherserver/index.js. Rebuild withnpx @anthropic-ai/mcpb pack mcpb snipperapp-<version>.mcpb.server.json— the entry for the official MCP registry (mcp-publisher publish).- The launcher and these docs are MIT licensed; SnipperApp 3 and its bundled
snipper-mcpbinary are proprietary.
Support
[email protected] — bug reports, feature requests and privacy questions. Documentation: https://snipperapp.com/docs/mcp-integration
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.