Granite

Built-in MCP server

Plug your knowledge base into any AI.

Granite speaks the Model Context Protocol natively. One remote URL for Claude, ChatGPT, Cursor, and your agents, or a local server that works offline against folders on your computer. You decide which vaults, and whether AI can write.

In short

The Model Context Protocol (MCP) is an open standard that lets AI assistants call tools and read data from other apps. Granite includes an MCP server, so AI tools can list, search, read, and (if you allow it) create and update your notes. The remote server is at https://mcp.granite.md/mcp; the local server runs with granite mcp and works offline.

Remote · Pro and Business

For Claude, ChatGPT, and teams

Streamable HTTP at one URL. Sign in with OAuth and approve the vaults and scopes, or use a scoped API token. Works with synced vaults and the knowledge base.

# Claude: Settings → Connectors → Add custom connector
URL  https://mcp.granite.md/mcp

# Claude Code
$ claude mcp add --transport http granite \
    https://mcp.granite.md/mcp

# Cursor · ~/.cursor/mcp.json
{ "mcpServers": { "granite": { "url": "https://mcp.granite.md/mcp" } } }

Local · Every plan, free included

For your own computer

A stdio server in the desktop app and CLI that reads a folder of notes directly, offline.

$ granite login
$ granite sync -watch Notes ~/Notes
$ granite mcp --vault ~/Notes
# add --read-only to offer only reading tools

Tools your AI gets

Everything an agent needs to work with notes.

MCP toolWhat it does
search_knowledge_baseRanked passages with citations from the knowledge base, filtered by permissions (remote)
search_notesSearch notes and return snippets with paths
read_noteFull content, frontmatter, and backlinks of a note
list_vaults, list_notes, list_tagsBrowse vaults, folders, tags, and recent changes
get_backlinks, get_outgoing_linksWalk the graph of linked notes
create_note, update_note, patch_noteWrite notes, only with write scope; updates are conflict-safe and versioned
move_note, delete_noteFile operations; deletes are soft and restorable

Safe by default

An MCP server your security team can approve.

Scoped tokens

Approve read, write, or knowledge base search, for chosen vaults only. Revoke any connection or token instantly.

Workspace AI policy

Admins decide whether AI tools may connect, which vaults they reach, and whether they can write.

Same permissions as people

An AI tool acting for a person never sees more than that person can open.

Audited and rate-limited

Connections and access are logged; rate limits protect your workspace.

FAQ

MCP questions.

New to MCP? Read what MCP is and how it works.

What is an MCP server for notes?

An MCP server for notes lets AI assistants that support the Model Context Protocol search, read, and edit your notes directly, instead of you copying and pasting. Granite's MCP server is built in: no plugins, scripts, or self-hosting needed.

How do I connect Claude to my notes?

In Claude, open Settings, then Connectors, choose Add custom connector, and enter https://mcp.granite.md/mcp. Sign in to Granite in the window that opens, choose what Claude may do and which vaults it may use, and approve. In Claude Code, run claude mcp add --transport http granite https://mcp.granite.md/mcp. For folders on your computer, use the local server from the desktop app or the granite command.

Can ChatGPT use my Granite knowledge base?

Yes. ChatGPT can reach Granite through MCP connectors, using the same remote MCP server and permissions as other AI tools.

Does the local MCP server need the internet?

No. The local server reads Markdown files in a folder on your computer and works offline. It's included on every plan, including Free.

Can an AI tool delete my notes?

Only if you give it write access. Write tools are off unless a token has write scope and the workspace policy allows it, deletes are soft, and every change is a version you can restore.

Give your company a memory it can trust

Free for one person. Business is free for 14 days, no card required.

  • Encrypted with a key per vault
  • Never used to train AI
  • Export as Markdown anytime