Set up a knowledge graph for Cursor

Cursor is quick at finding code that looks relevant to your prompt. What its agent lacks is typed structure: which function calls which, what imports what, how far a change propagates. Graphify builds that structure as a local knowledge graph in graphify-out/ and wires Cursor to it with a single always-on rule, so the agent queries the graph before it greps. (Comparing the two approaches head-on? See Graphify vs Cursor.)

The rule Graphify writes into Cursor

graphify cursor install creates exactly one file: .cursor/rules/graphify.mdc, with alwaysApply: truein its frontmatter, so Cursor includes it in every conversation. No hooks and no background process; Cursor does not expose PreToolUse-style hooks, so an always-on rule is the native mechanism (this is also why Cursor's setup differs from Claude Code's).

---description: graphify knowledge graph contextalwaysApply: true--- This project has a graphify knowledge graph at graphify-out/. **MANDATORY: Before using Read, Grep, Glob, or Bash to explore thecodebase, you MUST run graphify first:**

Below that, the rule names the three commands the agent should reach for: graphify query "<question>" for a scoped subgraph, graphify path "<A>" "<B>" for the dependency path between two symbols, and graphify explain "<concept>" for everything connected to a node. It also tells the agent to pass the same guidance to every subagent it spawns, to prefer the generated wiki when one exists, and to run graphify update . after modifying files. The file is wholly owned by Graphify (upgrades overwrite it, so stale wording never lingers) and it is committable, so one install covers the whole team.

Set it up

1. Install the CLI. If graphifyis not found in Cursor's terminal afterwards, run uv tool update-shell and restart Cursor.
2. From the project root: graphify cursor install.
3. Build the graph with /graphify . if you have not already; the install reminds you to. It produces graphify-out/graph.json plus the interactive graph.html and GRAPH_REPORT.md.
4. Ask Cursor a structural question and watch the agent query the graph first.

$ uv tool install graphifyy$ graphify cursor install

What a query looks like inside Cursor

Say you ask how the request handler reaches ModelField in a FastAPI-sized codebase. Instead of opening files, the agent runs a path query in the terminal and gets the route hop by hop:

$ graphify path "FastAPI" "ModelField"Shortest path (3 hops):  FastAPI --uses--> DefaultPlaceholder <--references-- get_request_handler() --references--> ModelField

For broad questions, graphify queryreturns a scoped subgraph under a token budget; for one symbol's full neighborhood, graphify explain lists every edge. Results carry file:line citations, and each edge is tagged EXTRACTED (parsed from the AST), INFERRED (a model connected it), or AMBIGUOUS (evidence, not fully resolved), so the agent knows what was found versus guessed. The full command set is in the CLI reference.

Optional: the MCP server in .cursor/mcp.json

The rule plus the CLI is the whole default setup. If you would rather expose the graph to Cursor as native MCP tools, register the server in your project's .cursor/mcp.json (or ~/.cursor/mcp.json globally):

{  "mcpServers": {    "graphify": {      "command": "python",      "args": ["-m", "graphify.serve", "graphify-out/graph.json"]    }  }}

That serves ten tools over the graph (query_graph, shortest_path, god_nodes, graph_stats, the PR-triage set, and more), documented in the MCP tools reference. The stdio server ships as an extra: uv tool install "graphifyy[mcp]".

Keeping the graph current

After edits, graphify update .re-extracts only the files that changed (AST-only, no API cost); the rule tells Cursor's agent to run it itself after modifying code. For hands-off freshness, graphify hook install adds post-commit and post-checkout git hooks that refresh the graph as you work.

Troubleshooting

• The agent ignores the graph: confirm .cursor/rules/graphify.mdc exists and its frontmatter says alwaysApply: true, then start a new conversation so the rule is loaded.
• The agent runs graphify but gets command not found: Cursor's terminal does not have the uv or pipx bin dir on PATH. Run uv tool update-shell (or pipx ensurepath) and restart Cursor.
• ModuleNotFoundError: No module named 'graphify': usually a plain pip install on Mac or Windows pointing at a different Python than the one recorded in graphify-out/.graphify_python. Install with uv or pipx instead.
• The agent still greps: that is by design when graphify-out/graph.json does not exist yet. The rule permits raw reads in that case, so build the graph first.

Cursor FAQ

Next steps

The quickstart walks the whole flow from install to first query, and the CLI reference covers every command the rule points Cursor at. Running Claude Code alongside Cursor? The same graph serves both: see Graphify + Claude Code, or browse all 17 assistants on the integrations page.