Skip to main content
Glama

Saveback

Saveback keeps the posts you save on every platform in one searchable local store (SQLite) and lets Claude search it over MCP. Once a week Claude can run the "you saved this and never opened it" digest through list_unseen, so saves stop disappearing into each platform's bookmarks.

The server stores items, serves them over MCP, and can fetch a public web page or YouTube video you give it (save_url). A Chrome extension captures saves from X, Instagram, LinkedIn, Facebook, Threads and YouTube.

Status: the extension exists and its adapters were built from the real page structure captured on 2026-11-13, but an end-to-end capture on live pages has not been verified yet. See docs/platform-status.md.

Quick start (local)

The package is not published to npm yet, so run it from a checkout (Node 22.13 or newer):

npm install
node server/cli.js

Register it with Claude Code, using the absolute path to your checkout:

claude mcp add saveback -- node /absolute/path/to/server/cli.js

After it is published this becomes claude mcp add saveback -- npx saveback.

Local mode serves MCP over stdio and a receiver for the extension on 127.0.0.1:7777 (override with SAVEBACK_PORT; the extension only reaches http://127.0.0.1:7777, so keep port 7777 if you use it in local mode). Data lives in ~/.saveback/saveback.db unless SAVEBACK_DB points elsewhere.

Related MCP server: MemoryAI Local

Browser extension

  1. In Chrome open chrome://extensions, turn on Developer mode, click Load unpacked and pick the extension/ folder of this checkout.

  2. Click the Saveback icon. The endpoint defaults to http://127.0.0.1:7777, which is where local mode listens (start it with node server/cli.js). For a hosted server paste its https:// URL and the token; Chrome asks permission for that origin the first time you click Capture. Other http:// endpoints are not accepted.

  3. Open the platform's saved page (table below) and click Capture. The page scrolls by itself while the extension reads it, so keep the tab open and visible until it finishes.

Platform

Open this page

x

x.com/i/history, Bookmarks tab

instagram

Saved > All posts, instagram.com/<you>/saved/all-posts/

linkedin

linkedin.com/my-items/saved-posts/

facebook

facebook.com/saved

threads

threads.com/saved

youtube

a playlist page such as Watch Later, youtube.com/playlist?list=WL

The result line in the popup reads:

  • N found, M saved: N posts were read from the page and M were stored. Posts you already saved are updated, not duplicated. (K rejected by the server) means the server refused K of them as invalid.

  • 0 found, selectors may be outdated: the page matched but no posts were read, so the platform probably changed its markup. See docs/platform-status.md.

  • This is not a saved page. Open ...: the current tab is not the page for that platform.

  • Cannot reach Saveback at <endpoint> or Check your token.: the server is not running or the token is wrong. Items not sent are counted (N found, M saved, K not sent); click Capture again to retry.

Capture stops after 5 scroll steps with nothing new, or at 2000 items. Per-platform limits are in docs/platform-status.md; to add a platform see docs/adding-a-platform.md.

MCP tools

Tool

What it does

search_saved

Full-text search over text, author, url, transcript and notes; optional platform and date filters

get_saved

One item in full, including the whole transcript

list_recent

Newest saves first

list_unseen

Saves not yet dealt with; older_than_days resurfaces forgotten ones (the digest)

mark_seen

Mark an item as dealt with

add_note

Attach a searchable note to an item

tag

Replace the tags on an item

save_url

Fetch a public web page or YouTube video and store it; attempts to fetch the transcript (YouTube may not serve captions to non-browser clients, so it can be empty)

Self-host on Railway

Hosted mode runs the same server behind a bearer token so the extension and Claude can reach it from anywhere. The repo ships a railway.json (start command node server/cli.js, health check /health).

A first deploy fails its health check until SAVEBACK_MODE=remote and SAVEBACK_TOKEN are set (without SAVEBACK_MODE=remote the server starts in local mode on 127.0.0.1 and Railway cannot reach it). Set the volume and SAVEBACK_DB before the first deploy too: without them the server still starts, but stores data on ephemeral disk and loses it on redeploy.

  1. Create a Railway project from this repo.

  2. Add a volume mounted at /data.

  3. Set the variables:

    • SAVEBACK_MODE=remote

    • SAVEBACK_TOKEN a long random secret, e.g. from openssl rand -hex 32 (the server trims it and refuses to start without it or with fewer than 16 characters)

    • SAVEBACK_DB=/data/saveback.db

    • RAILPACK_NODE_VERSION=22 to make sure the build uses Node 22 (the server needs 22.13 or newer; Railway's builder also reads engines.node from package.json, this variable just removes any doubt)

  4. Generate a public domain. Railway sets PORT itself; the server listens on it.

  5. Check it: curl https://<your-domain>/health returns 200 without a token. /ingest and /mcp need Authorization: Bearer <token>.

  6. Connect Claude:

claude mcp add --transport http saveback https://<your-domain>/mcp --header "Authorization: Bearer <token>"

claude mcp add --header stores the token in Claude's config (~/.claude.json). To rotate it, change SAVEBACK_TOKEN on the server, then run claude mcp remove saveback (add -s user or -s project if you added it with that scope) and add the server again with the new token (adding it twice under the same name fails), and paste the new token into the extension popup.

Privacy and safety

  • Your data stays in your own SQLite file (local disk, or the volume on your own Railway deploy). Nothing is sent to a Saveback service.

  • The extension reads a page only when you click Capture: the popup injects the capture script into the active tab at that moment, and there are no content scripts that run on their own.

  • Capture happens in your own logged-in browser. Scraping logged-in pages may go against a platform's terms of service; use at your own risk.

  • Saved content is third-party text and can contain instructions aimed at Claude. Do not auto-approve save_url (it fetches whatever URL it is given), and treat saved text as data, not commands.

  • Hosted mode is single-tenant: one deploy, one token, one person's data. save_url refuses private and local addresses (a small DNS-rebinding window is a known limit).

License

MIT, see LICENSE.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Turns likes, bookmarks, and saved links into persistent project-aware memory, enabling AI agents to search, import, and process personal content through a local SQLite-backed MCP server.
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Provides persistent long-term memory and local RAG document search for Claude Desktop, fully offline. Enables memory save/search and document retrieval via MCP.
      MIT
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables users to create, link, explore, and synthesize atomic notes through Claude and other MCP-compatible clients, with semantic links, tags, search, and a Markdown-backed SQLite index.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables users to create, link, explore, search, and synthesize atomic Markdown notes through Claude and other MCP-compatible clients, with tagging, semantic link types, and dual Markdown/SQLite storage.
      -