/cortexdb-graph

description: Open the brain's knowledge graph — a live 3D view, or a static file to send

Two views of the same graph. Pick by what the user asked for, and default to the live one.

Live 3D (default)

Ask for it with the **`serve_graph_3d`** MCP tool. It opens a rotatable, glowing 3D view in the browser and returns its URL. Prefer this whenever the user says show / see / open / watch the graph, or mentions 3D, live, or rotating. When they ask for their memories as a library, books or shelves, pass `view: "library"`: the same server draws each memory as a book on its project's shelf.

It is served from inside this MCP server, which is what makes it live: nodes and relations appear as they are written, and **every query, save and relation this server handles lights up the nodes it touched** — so the user watching the page sees the brain react to what you do next. Calling the tool again returns the same URL rather than opening a second view.

After it opens, tell the user the URL, and that the view keeps updating on its own — they do not need to re-run anything. Worth mentioning once: **Trace path** picks two nodes and lights the chain between them, **Orbit** turns the scene, and dragging takes the camera back.

If the user wants it from a terminal instead, the binary serves it too — but a command-line view is its own process, so it sees the graph change and *not* the queries:

bin=$(ls -t ~/.claude/plugins/data/cortexdb-cortexdb/bin/cortexdb-mcp-* 2>/dev/null | head -1)
[ -n "$bin" ] || bin="${CORTEXDB_MCP_BIN:-cortexdb-mcp}"
"$bin" --graph-3d          # prints the URL, opens a browser, runs until Ctrl-C

Static file (when it has to be sent)

When the user wants something they can **attach to a message, keep, or open without this session running**, render the self-contained file instead — the live view is a server that dies with this process, so it cannot be sent anywhere.

bin=$(ls -t ~/.claude/plugins/data/cortexdb-cortexdb/bin/cortexdb-mcp-* 2>/dev/null | head -1)
[ -n "$bin" ] || bin="${CORTEXDB_MCP_BIN:-cortexdb-mcp}"
html=$("$bin" --graph-html)        # writes to ~/.cortexdb/graph/ and prints the path
echo "$html"
( command -v open >/dev/null && open "$html" ) || ( command -v xdg-open >/dev/null && xdg-open "$html" ) || true

Add `--organize` only if asked: it extracts entities from stored text with no LLM, and it **writes to the graph**. It is ignored against a shared brain.

If the graph looks thin

Say why rather than leaving them to wonder: the graph grows from knowledge saved with entities and relations (`knowledge_save`, `upsert_relations`) or ingested through GraphRAG. Free-text memories with nothing linked do not add nodes.

$ARGUMENTS