Graph memory Grounded answers

The knowledge graph your AI can reason over.

Open-source and on-device. One command maps your repo into a graph your AI assistant traverses instead of grepping. Every answer traces to a path you can audit.

Get started Star on GitHub★ 0

Talk to the founder Talk to the founder

Backed byOn-device · no telemetry

Enterprise · early access

Be first when Graphify Enterprise ships.

86,755GitHub stars3,245 to go → 90,000

3M+PyPI downloads[EXTRACTED]

17AI assistants[EXTRACTED]

Already building on Graphify[INFERRED]

MIT license · 36 languages

How it works

Install → graph → query.

Step 01

Run `/graphify .` in the assistant you already use

Graphify installs as a skill. No extension, no account. Pick your assistant for the exact commands.

What are you coding with?

$ uv tool install graphifyy
$ graphify install  # add the skill
> /graphify .  # inside Claude Code

package is graphifyy— yes, two y's

Then star it on GitHub · join the Discord

Step 02

Get the graph

One run maps the repo into three local files your assistant reads from. graph.html, GRAPH_REPORT.md, and graph.json.

Step 03

Query, don't grep

$graphify query "who owns billing?"
→ 1 path · 2 hops · EXTRACTED

Your assistant follows real paths through the graph with query, path, and explain.

The payoff

The answer is a path, not a vibe

Your assistant cites the edges it walked — each one tagged with where it came from, so you can check it instead of trusting it.

billing_servicestripe_clientpayments-team[EXTRACTED]

See it

Your whole codebase, as one graph.

graph.html — fastapi · graphify

graphify's interactive graph.html rendering the FastAPI repository: a dark force-directed knowledge graph of colored symbol nodes and relationship edges, with a Communities sidebar listing auto-detected clusters — APIRouter (96 nodes), SecurityBase (71), v2 (56), FastAPI (46) — each with a checkbox filter and color swatch.

god node

The most-connected symbol in the repo. Bigger dot, more dependents; a change here ripples wide.

one color, one community

graphify clusters tightly linked code into modules automatically. This orange region is one of them.

filter by community

Every detected cluster, named and counted. Toggle one on to isolate a module.

how to read itnode = one symbol (function, class, file)color = auto-detected community (a module)big node = god node (most connected)line = an import or call

1god node. The most-connected symbol in the repo. Bigger dot, more dependents; a change here ripples wide.
2one color, one community. graphify clusters tightly linked code into modules automatically. This orange region is one of them.
3filter by community. Every detected cluster, named and counted. Toggle one on to isolate a module.

FastAPI's own repo, mapped by one run of /graphify . · communities, god nodes, search, click-to-inspect.

In the wild

In their words.

The adopters under the hero each link to public proof. This is one integration up close, plus results the community reported.

“The data already lives in Rootly. The graph makes the structure visible.”

Sylvain KalacheRootly AI LabsIncident management · SRE[EXTRACTED]

Rootly turned its incident data (incidents, alerts, teams, services) into a queryable knowledge graph with Graphify.

View the integration

Real results, reported by the community

71.5×fewer tokens in one user's Claude Code and Graphify setup.[EXTRACTED]lucasrosati/claude-code-memory-setup 79×fewer tokens on a 496K-token codebase, with zero vector database in the stack.[EXTRACTED]Steve Scargall, MemVerge

Try it on your codebase.

One command, about five minutes, entirely on-device.

Get started Read the docs →

Why a graph, not grep

Every answer traces to a real path.

Grep hits and fuzzy chunks make your assistant guess. A knowledge graph gives it structure to reason over.

Grep reads files and forgets them. Every session starts from zero.

The graph persists. Your assistant opens already knowing the architecture.

RAG retrieves fuzzy chunks and hopes the model guesses right.

Your assistant traverses connected entities, hop by hop.

Answers backed by an opaque similarity score.

Answers backed by a path you can audit, tagged extracted, inferred, or ambiguous.

GRAPH_REPORT.md

God nodes: The most-connected files in the repo, ranked. Change these carefully.
Surprising connections: Cross-module dependencies you didn't know you had, ranked by unexpectedness.
The “why”: NOTE, WHY, and HACK comments become nodes attached to the code they explain.

Side by side

Graph, chunks, or grep.

Full comparison →

GraphifyVector DB / RAGgrep

Structure

GraphifyA real graph — typed entities and relationships

Vector DB / RAGFuzzy chunks, ranked by similarity

grepRaw text matches, no connections

Provenance

GraphifyA path you can audit; every edge tagged extracted, inferred, or ambiguous

Vector DB / RAGAn opaque similarity score

grepThe matching file and line

On-device

GraphifyYes — parsing and storage are local, no telemetry

Vector DB / RAGVaries; embedding APIs are usually remote

grepYes

Works in any assistant

GraphifyOne skill installs in 17 assistants, plus an MCP server

Vector DB / RAGA custom pipeline per setup

grepBuilt into most assistants already

License

GraphifyMIT, open source

Vector DB / RAGMixed — open engines, proprietary hosted services

grepOpen source

Confidence

Every edge says how it knows.

Graphify tags each relationship it maps with the evidence behind it. When your assistant answers from the graph, you can tell what was read from source and what was reasoned.

one edge from graph.jsonpayments/billing.py · L41

billing_servicecallsExtractedstripe_client

Extracted

Found in the code. The relationship exists at a file and line you can open.

Inferred

Reasoned from structure and naming. Usually right, worth checking before you rely on it.

Ambiguous

The evidence points more than one way, and the graph tells you so.

The rest of the toolkit

graph.json: The whole repo as entities and relationships, parsed by 36 tree-sitter grammars and stored on disk.
/graphify .: Installs as a skill in 17 assistants, including Claude Code, Cursor, Copilot, Codex, Gemini CLI and Aider.
graphify prs: A PR dashboard in the terminal: CI status, AI triage, merge-conflict risk.
MCP server: Serves the graph over stdio on one machine, or over HTTP for the whole team.
On-device: Parsing runs locally, no telemetry. Bring your own model, or run Ollama.

Pick your door

Developers run it. The whole team reads it.

~/your-repozsh

$ /graphify .

✓ graph ready · MCP server live

$ graphify query "who owns billing?"

→ 1 path · 2 hops · EXTRACTED

For developers

Map your codebase in one command

Run /graphify . in your coding assistant, query with graphify query and explain, and drop the MCP server into Claude Code, Cursor, and 15 more.

Understand a new codebase. Query god nodes instead of reading 100 files.
Review PRs faster. graphify prs triages CI, reviews, and merge-conflict risk.

Install in one command

For teams

The graph is a file your team can share

Graphify writes graph.json into the repo. Commit it, or serve it over HTTP with graphify.serve. A new hire asks "who owns billing?" and gets a two-hop path through the actual code, not a week-old Slack thread.

Onboard engineers. NOTE, WHY, and HACK comments become linked nodes to follow.
Ground your AI assistant. Answers come from real paths, tagged extracted or inferred.

Set it up for your team

Trust

Your code never leaves your machine.

Every hosted indexer asks you to ship your repo to someone else's cloud first. Graphify doesn't, because it can't — there is no server in the loop.

On-device

Parsing runs locally — 36 tree-sitter grammars over your repo, on your hardware. The graph is a file on your disk, not rows in a hosted index.

No telemetry

No usage pings, no crash reports, no analytics. There's nothing to opt out of, because nothing is sent.

MIT, auditable

The entire source is MIT-licensed on GitHub. You don't have to trust this page — read what the code does.

Read the source

Self-host the MCP server

Serve the graph over stdio on one machine, or over HTTP on your own infrastructure for the team. We host nothing.

The full picture: your code stays local. The only thing that leaves your machine is the queries your assistant already sends to your own model provider, under your own keys — Graphify adds no new destination.

How we think about security →Graphify for enterprises →

Pricing

$0. MIT. Free forever.

Everything on this page is the open-source product. It runs on your machine: no limits, no account, no card.

Install the CLI Pricing details →

Enterprise features (verification, code review, CTO digest, Jira) are in early access. Join the list

Start in one command

Try it on your own repo.

Get started

86,755 stars on GitHub. Read the code.

Evaluating for a team? See Graphify Enterprise →

The knowledge graph your AI can reason over.

Install → graph → query.

Run /graphify . in the assistant you already use

Get the graph

Query, don't grep

The answer is a path, not a vibe

Your whole codebase, as one graph.

In their words.

Try it on your codebase.

Every answer traces to a real path.

Graph, chunks, or grep.

Every edge says how it knows.

The rest of the toolkit

Developers run it. The whole team reads it.

Map your codebase in one command

The graph is a file your team can share

Your code never leaves your machine.

On-device

No telemetry

MIT, auditable

Self-host the MCP server

$0. MIT. Free forever.

Try it on your own repo.

Run `/graphify .` in the assistant you already use