Savebase MCP Server: Connect Bookmarks to ChatGPT and Claude
Your saved research can become useful context inside an AI conversation. The Savebase MCP server lets compatible assistants search your bookmarks, read notes and retrieve available video transcripts without repeatedly copying links into a chat.
If your data is already in Savebase, you can use its hosted connection. If you need to query files or a database on your computer, a custom MCP server serves a different purpose. This guide explains the choice and shows how to connect Savebase.
What is an MCP server?
MCP stands for Model Context Protocol. It gives an AI application a standard way to discover tools and request information from another service. A server defines the available tools; the assistant calls them to retrieve context for your request.
Read the official MCP architecture overview.
For example, an assistant can search a research folder, retrieve selected bookmarks and compare their contents. MCP is the connection mechanism. It does not guarantee an accurate answer, unrestricted compatibility or private processing by every AI provider.
Savebase MCP versus a custom server for local data
- Use Savebase MCP for content already saved in your Savebase account. It is a hosted HTTP connection; you do not install a local Savebase server or maintain an SDK to connect your library.
- Build or configure a local MCP server for folders, Markdown files, SQLite databases or internal services that Savebase does not expose. Savebase MCP does not scan your hard drive or provide arbitrary database access.
Running a local server does not guarantee that retrieved text stays only on your device. A local server can send tool results to a cloud assistant. An entirely local workflow also requires a local model and careful control of the client, networking and logs.
Before connecting Savebase
You need a Savebase account whose current entitlement includes MCP access, plus ChatGPT with the Savebase plugin or another compatible MCP client. Client availability and workspace restrictions can vary. Check your Savebase settings for access.
See the Savebase MCP setup documentation.
How to connect Savebase to ChatGPT
1. Open the published Savebase plugin
Open the official Savebase plugin listing in ChatGPT and install it using the add control. This is the published-plugin route; you do not need to create a custom integration in Developer mode.
Open the Savebase plugin in ChatGPT.
2. Sign in and review access
When prompted, sign in to your Savebase account. Review the consent screen and approve read-only access if you want to connect this account. The assistant does not receive your Savebase password.
3. Choose Savebase in a conversation
Start a new ChatGPT Work conversation. Choose Savebase from the composer or mention @Savebase, then ask it to find a specific saved topic. Request source links so you can check the result.
Connect Savebase to Claude and other MCP clients
In Claude, open Settings → Connectors, add a custom connector named Savebase and enter the server URL below. Add the connector, connect it and complete the Savebase browser sign-in. Organization policies may control which connectors you can use.
https://mcp.savebase.space/mcp
Check Claude’s current custom connector instructions.
For Claude Code, add the HTTP server with the command below. The user scope makes the configuration available across your projects. Start Claude Code and use /mcp to complete authentication or inspect the connection.
claude mcp add --transport http --scope user savebase https://mcp.savebase.space/mcp
Read the Claude Code MCP instructions.
Cursor, VS Code and other compatible clients use the same server URL, with their own HTTP MCP configuration and sign-in flow. Prefer browser authentication. Manual tokens are an advanced fallback for older clients; keep them private and revoke unused tokens.
What the nine Savebase read tools can do
search_bookmarks: Search keywords across titles, saved content, notes, creators, URLs and existing transcripts. Narrow results by platform, folder, tags, creator or saved date.
list_bookmarks: Browse saved items without a keyword, with optional platform, folder, tag and date filters.
list_recent_bookmarks: Find items saved during named periods, such as today, this week or the last 30 days.
get_bookmark: Retrieve one bookmark’s saved details, notes, media links, folders and tags. Long content may be shortened; transcripts start with a preview.
get_bookmark_transcript: Read an existing video transcript in chunks. This retrieves a transcript already available in Savebase; it does not transcribe a video.
find_bookmark_by_url: Check whether a link is already saved, with normalization for tracking parameters and a trailing slash.
list_folders: Discover folder names, IDs and nesting before narrowing a request to a folder.
list_tags: Discover existing tag names before filtering bookmarks by tag.
get_library_stats: Summarize platforms, recent activity, folders, tags and transcribed videos. Statistics can be approximate for a large library.
These tools read your library. They do not add bookmarks, change notes, move items, assign tags or delete content. Listing tags or folders gives the assistant information; it does not grant permission to reorganize them.
Useful prompts for saved research
- Find my saved posts about OAuth in the last 30 days. Show the source links and tell me which results contain my own notes.
- List my folders, then find articles about pricing in my Product Research folder. Compare the retrieved sources and separate direct quotations from analysis.
- Search my existing video transcripts for rate limits. Read the relevant transcript chunks before summarizing; say if no transcript is available.
- Have I already saved this URL? Return the matching bookmark and its source link.
Search is based on keywords and filters, not a promise of semantic vector retrieval. Try alternative terms if a concept has different wording in your saved material. Ask the assistant to retrieve the actual bookmark before relying on a search snippet.
Results are paginated and some fields are shortened to fit the conversation. For a complete review, ask the assistant to follow additional pages and transcript chunks. A single page of results is not evidence that it has reviewed your entire library.
Privacy, permissions and disconnecting
Savebase authenticates the connection and restricts requests to the connected account. Browser authorization currently grants the bookmarks.read scope. Read-only access prevents library changes, but retrieved information still reaches the assistant handling your request.
That information can include saved text, notes, source and media URLs, folder names, tags and requested transcripts. Review the AI provider’s privacy and data-use settings before connecting sensitive research. MCP itself is not a guarantee about training or retention.
Revoke a connection in Savebase under Settings → MCP Server → Connected apps → Disconnect. Revoke manual tokens separately in the advanced token controls. Disconnecting stops future authorized access; it does not erase content already present in an AI conversation.
Read Savebase’s privacy policy.
Fix connection and retrieval problems
- No sign-in prompt: confirm the server address ends in /mcp and that the client supports remote HTTP MCP and browser authentication.
- Unauthorized or expired connection: reconnect and sign in again. A 403 access error can mean the Savebase account does not currently include MCP access.
- No results: confirm the item exists in the connected account. Check folder names and tag names, broaden the saved-date filter and try a shorter keyword.
- Missing video content: verify that a transcript exists in Savebase. This connection retrieves existing transcripts and cannot create a missing one.
When to build a custom MCP server instead
For local data retrieval, start by choosing a small, explicit data source and a narrow read tool. For example, expose keyword search over one approved documentation folder and a separate read tool that accepts only IDs returned by that search.
- Use a maintained MCP SDK for your programming language. Follow its current setup instructions instead of copying old dependency versions from a tutorial.
- Validate tool inputs, restrict file paths to the approved directory and use read-only database access. Add result limits and pagination so one request cannot return the whole archive.
- Choose a transport your target client supports. Local stdio and hosted HTTP have different deployment and authentication requirements; OAuth is not exclusive to either data location.
- Test discovery and retrieval in an MCP client or Inspector. Check invalid paths, empty results, authorization failures and oversized responses before connecting private data.
Keep credentials out of tool output and treat retrieved documents as untrusted content. Their text should be evidence for an answer, not authority to run commands or reveal other data. Local storage and a cloud assistant require an explicit data-sharing decision.
If the material you need is already in Savebase, start with its existing connection. For data that must remain on your own infrastructure, design the server and model workflow around that requirement before importing or exposing it.
Frequently asked questions
Is Savebase MCP a local file server?
No. Savebase MCP is a hosted connection to your saved Savebase library. It does not search arbitrary files on your computer. Use a separate local server for local folders or databases.
Can ChatGPT change my Savebase bookmarks?
The current Savebase connection is read-only. Its tools can search and retrieve saved information, but cannot add, edit, move or delete bookmarks.
Do I need an API key to connect Savebase?
For the published ChatGPT plugin and clients supporting Savebase’s browser authentication, sign in with your Savebase account. You do not need to copy an API key. Older clients may require an advanced manual token.
Can MCP search inside saved videos?
Savebase can search an existing transcript and return transcript chunks. A video without an available transcript cannot be searched by its spoken content through these tools.
Does running MCP locally guarantee privacy?
No. Retrieved text may still be sent to a cloud model by your AI client. Check the full data path, provider settings and logging; running the server locally alone does not make processing entirely local.