对话与笔记保存MCP服务器

The Dump MCP Server

An MCP (Model Context Protocol) server that lets LLM clients (Claude Desktop, Claude Code, Cursor, etc.) save conversations directly to The Dump — and read your saved notes back as context, so your assistant can answer questions about them.

Full setup guide: https://thedump.ai/mcp

Hosted server (recommended)

The same tools are available as a hosted MCP server at https://mcp.thedump.ai/mcp — no install, no password in your chat tool. Add it as a custom connector / remote MCP server and sign in with your The Dump account in the browser:

• claude.ai / Claude Desktop: Customize → Connectors → Add custom connector → URL https://mcp.thedump.ai/mcp (leave the OAuth client on "Use Claude's published identity").
• Claude Code: claude mcp add --transport http the-dump https://mcp.thedump.ai/mcp
• Cursor / others: add the URL as a Streamable HTTP server; it supports Dynamic Client Registration.

The hosted server differs from the local one in two ways. It cannot read files on your machine, so create_note's file_path is unavailable there (use content or a public file_url). And because you sign in through the browser, it has no login, signup or logout tools. Everything below describes the local (stdio) server, which keeps working as before.

Quick start (npm)

No install step needed — point your MCP client at npx:

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "the-dump": {
      "command": "npx",
      "args": ["-y", "the-dump-mcp"]
    }
  }
}

Claude Code

claude mcp add the-dump -- npx -y the-dump-mcp

Cursor

Add the same JSON block as Claude Desktop to .cursor/mcp.json in your project or ~/.cursor/mcp.json globally.

Running from source (alternative)

The source on main is version 2.0.0, which adds the hosted server; the npm package stays on 1.5.0. The hosted (HTTP) entry is dist/http.js; it needs MCP_PUBLIC_URL, WORKOS_AUTHKIT_DOMAIN and THE_DUMP_BASE_URL (plus a Google service identity to call The Dump's API) and is deployed to Cloud Run by the monorepo's CD. npm test runs its contract tests.

git clone https://github.com/George-Labs-AI/The_Dump_MCP.git
cd The_Dump_MCP
npm install
npm run build

Then use "command": "node", "args": ["/absolute/path/to/The_Dump_MCP/dist/index.js"] in the configs above (or claude mcp add the-dump -- node /absolute/path/to/The_Dump_MCP/dist/index.js).

Authentication

No tokens or environment variables needed. Once the server is connected, tell your assistant to "log in to The Dump" — it will use the login tool with your Dump email and password. Credentials are stored locally in ~/.the-dump/credentials.json (mode 0600) and refresh automatically; use the logout tool to clear them.

Tools

Saving

The share_* tools save a conversation transcript (with role labels and a "Saved from …" provenance line). create_note is different: it saves content exactly as given, with no conversation framing, and can attach a file — so a grocery list, a draft, an HTML artifact, or a photo lands in The Dump the same way it would from the iOS app. Pass exactly one of content, file_path (absolute or ~/… on this machine), or file_url (public http/https link). Images are described and OCR'd, audio is transcribed, and documents are parsed, then The Dump chooses the category, title, and type. Supported types: jpg/png/gif/bmp/webp/heic/tiff, pdf/doc/docx/txt/md/html/json/xml/csv, and mp3/wav/flac/m4a/aac/ogg/webm/opus, up to 100 MB. For a text note, title is prepended as a heading; filename sets the extension (e.g. report.html).

Reading your notes

Ask things like "summarize my notes in the Recipes category from last month" — your assistant will translate that into the right search and filters. You can only ever read notes belonging to the account you're logged in as; the server derives your identity from your auth token, never from request parameters.

Routines (read-only)

Routines are long-running processes that maintain living "canon" documents (e.g. a project dashboard or plan) from your notes, and queue judgment calls as approval requests (ASKs).

Ask things like "what does my renovation plan say about the septic system?" — your assistant will find the right routine and read its canon docs. These tools are strictly read-only: canon documents are written only by the routine's runner, and approval requests are answered in The Dump's web UI.

Very large canon documents are returned as a short preview by default; your assistant will ask you before loading the whole thing into the conversation (some canon docs are hundreds of thousands of characters).

A note on agent safety: retrieved notes are returned clearly framed as data, with instructions to the model not to treat note content as commands. Still, notes can contain text you saved from elsewhere (web clippings, OCR'd images, shared conversations). If you run an agent with broad, auto-approved permissions over your notes, you are trusting everything you've ever saved — keep permission prompts on when in doubt. The same applies to create_note with file_path: it uploads any file the assistant can read on this machine into your own account, so keep the permission prompt on for it.

Tasks

The Dump automatically extracts action items from each new note — a To-Do list (kind: do) and a To-Buy list (kind: buy). Tasks live in their own table; checking one off never changes the note it came from.

Ask things like "what's on my shopping list?" or "mark the dimmer-switch task done". Task text is model-generated from your own notes and is returned framed as data, not instructions.

Account (local server only)

The hosted server signs you in through the browser instead, so it does not offer these.

Routine status and current coverage

list_routines renders the server's per-routine runner status: the verdict from the newest shift (OK / PARTIAL / FAILED / NOTHING_TO_DO / UNKNOWN, or SILENT / PAUSED when the server derives them), the last shift, reasons, and the next expected report. Some routines also carry a newer six-stage status.workflow coverage block in the REST API and on the web Routines page; this tool does not yet render it. A FAILED verdict here describes the most recent shift and does not by itself mean the routine's current work is pending.