graph-autotagger-mcp

A FastMCP server that reads a pre-computed Obsidian knowledge graph and suggests wikilinks for your notes. Designed as a companion to graphify — run graphify to build the graph, then this MCP surfaces connections as you write.

Tools

Tool	What it does
`suggest_links`	Suggest wikilinks for a note based on token overlap with graph node labels
`check_god_nodes`	Check which highly-connected nodes (top 20 by degree) are relevant to the note
`community_match`	Find which graph communities the note most closely belongs to
`log_suggestion_decision`	Log accepted/rejected/ignored decisions to a local SQLite database

Install

Open Claude Code, paste:

/plugin marketplace add adelaidasofia/graph-autotagger-mcp
/plugin install graph-autotagger-mcp@graph-autotagger-mcp

After install, set GRAPH_JSON_PATH (see Environment variables below) and restart Claude Code, then ask:

"Suggest wikilinks for this note: [paste note content]"

Legacy install

pip install fastmcp

Clone:

git clone https://github.com/adelaidasofia/graph-autotagger-mcp.git
cd graph-autotagger-mcp

Set the path to your graph.json:
```
export GRAPH_JSON_PATH="~/vault/.graph/graph.json"
```
Or pass it at registration time (see below).
Self-test:
```
python3 server.py
```

claude mcp add graph-autotagger -s user -- \
  env GRAPH_JSON_PATH="$HOME/vault/.graph/graph.json" \
  python3 /path/to/graph-autotagger-mcp/server.py

Restart Claude Code, then ask:

"Suggest wikilinks for this note: [paste note content]"

Environment variables

Variable	Default	Description
`GRAPH_JSON_PATH`	`~/vault/.graph/graph.json`	Path to your graphify output
`AUTOTAGGER_DB`	`~/.config/graph-autotagger/log.db`	SQLite log for suggestion decisions

How it works

The server loads graph.json once at startup and caches it in memory. Node labels are tokenized and matched against note content using token overlap scoring. God Nodes (top 20 by degree) get special treatment since connecting to them creates high-value cross-links.

The log_suggestion_decision tool lets you build a feedback dataset over time: accepted suggestions can be used to tune the relevance threshold; rejected ones reveal graph noise.

graph.json format

Expects the output format from graphify (networkx node_link_data):

{
  "nodes": [{"id": "node-id", "label": "Node Label", "community": 0}],
  "links": [{"source": "node-a", "target": "node-b"}]
}

Note: networkx serializes edges under "links", not "edges".

Related MCPs

Same author, same architecture pattern (FastMCP, draft+confirm on writes where applicable, vault auto-export, MIT):

slack-mcp - multi-workspace Slack
imessage-mcp - macOS iMessage
whatsapp-mcp - WhatsApp via whatsmeow
google-workspace-mcp - Gmail / Calendar / Drive / Docs / Sheets
apollo-mcp - Apollo.io CRM + sequences
substack-mcp - Substack writing + analytics
luma-mcp - lu.ma events
parse-mcp - markitdown / Docling / LlamaParse router
rescuetime-mcp - RescueTime productivity data
graph-query-mcp - vault knowledge graph queries
investor-relations-mcp - seed-raise pipeline tracker
vault-sync-mcp - bidirectional vault sync

Telemetry

This plugin sends a single anonymous install signal to myceliumai.co the first time it loads in a Claude Code session on a given machine.

What is sent:

Plugin name (e.g. slack-mcp)
Plugin version (e.g. 0.1.0)

What is NOT sent:

No user identifiers, names, emails, tokens, or API keys
No file paths, message content, or anything from your work
No IP address is stored after dedup processing

Why: Helps the maintainer know which plugins people actually install, so attention goes to the ones that get used.

Opt out: Set the environment variable MYCELIUM_NO_PING=1 before launching Claude Code. The hook will skip the network call entirely. Already-pinged installs leave a sentinel at ~/.mycelium/onboarded-<plugin> — delete it if you want to reset state.

License

MIT

Built by Mycelium AI. Full install or team version at diazroa.com.

FastMCP server for Obsidian wikilink suggestions from a pre-computed knowledge g