MCP for AI agents
Looking for the CLI or skill instead? MCP is one of three agent surfaces — see Agent access: MCP, CLI & skill for the full comparison and setup for all three.
Fluffmind exposes vault tools over Model Context Protocol:
| Tool | Description |
|---|---|
get_workspace | Confirm bound workspace + token scope |
search_notes | Search by title or id |
read_note | Read markdown + frontmatter |
write_note | Create/update via writeToWorkspace (write scope) |
list_backlinks | Incoming wikilinks |
get_graph | Vault link graph |
create_task | Append - [ ] task (write scope; default note: inbox/tasks) |
Remote HTTP (recommended for staging / production)
- Sign in as a workspace owner.
- Open Settings → workspace → Agents.
- Enable agent access, create a named token (
readorwrite). - Copy the secret once and configure your client:
json
{
"mcpServers": {
"fluffmind": {
"url": "https://your-fluffmind.example.com/api/mcp",
"headers": {
"Authorization": "Bearer fm_agent_…"
}
}
}
}The Bearer token binds the agent to that workspace — the same token also works for the REST API and CLI. Revoke tokens from the same settings page. Disabling agent access rejects tokens even if they are still valid hashes. Legacy fm_mcp_… tokens issued before this rename still authenticate.
Session cookies still work as a fallback when no Bearer header is sent (browser).
Local (stdio)
For Claude Code, Cursor, or another stdio MCP client on the same machine:
sh
VAULT_PATH=/absolute/path/to/your/vault pnpm --filter @fluffmind/web mcp:stdioExample Cursor config (.cursor/mcp.json):
json
{
"mcpServers": {
"fluffmind": {
"command": "pnpm",
"args": ["--filter", "@fluffmind/web", "mcp:stdio"],
"env": {
"VAULT_PATH": "/absolute/path/to/your/vault"
}
}
}
}