Your assistant, with a memory of your codebase

You already have an AI assistant you like. The problem isn't the model — it's that every session starts from zero. Graphify is an open-source CLI and skill that maps your codebase into a knowledge graph on your machine, so the assistant queries structure instead of grepping and re-reading files it saw yesterday.

The problem: retrieval by grep

Ask "how does auth work here?" and most assistants answer by searching for strings, then reading whole files into context. You pay for those tokens every time, in every session, and the answer still misses anything that isn't a literal text match — the caller of a caller, the config that wires two modules together, the import three hops away. For a solo developer, that shows up as two costs: money burned on retrieval, and context window burned on files instead of the actual task.

OSS maintainers feel a second version of it: contributors (and their assistants) ask the same architectural questions over and over, and the mental map of the project lives only in your head.

How the graph helps

One command — /graphify . in your assistant, or graphify from the terminal — parses your code on-device with bundled tree-sitter grammars across 36 languages. Functions, classes, imports, and calls become nodes and edges straight from the AST, tagged EXTRACTED. Non-code context (docs, schemas, config) is connected by whichever model backend you already use, tagged INFERRED. Anything the parser couldn't resolve is kept and flagged AMBIGUOUS rather than silently guessed. The full pipeline is on the concepts page.

Then your assistant stops grepping. It runs graphify query for everything connected to a symbol, graphify pathfor the route between two points, and receives the relevant subgraph — a handful of nodes and edges, not the files that contain them. That's the token reduction mechanism: less context spent on retrieval, more room for the work. It installs as a skill or an MCP server in 17 assistants, including Claude Code, Cursor, Codex, and Copilot — the full list is on the integrations page.

For maintainers, the build step also emits an interactive graph.html and a GRAPH_REPORT.md naming the communities, god nodes, and surprising connections in your project — an architecture overview you can link from your README instead of re-explaining in every issue thread.

Why this fits a solo workflow

No account, no API key for parsing, no telemetry, no server. The graph is a file on your disk; the source is MIT-licensed and you can read exactly what it does. Install is one command — uv tool install graphifyy — and free means free: the CLI is the product, not a trial. If you've been tempted to wire up an embedding pipeline for codebase context, Graphify vs RAG covers why a graph keeps the structure that chunking throws away.

What it won't do

Honest limits: if your repo is small enough that your assistant can read it whole, a graph adds ceremony you don't need. Graphify doesn't make a weak model strong — it feeds a good one better context. And it doesn't change where your assistant sends its own prompts; it just adds no new place your code travels.

Try it on your own repo

The quickstarttakes about two minutes: install, build the graph, ask your assistant something that used to require reading five files. If the answer arrives as a path through your real code, you'll know it's working.