Draw the architecture.
Keep the code honest.
Graphlit turns a whiteboard sketch into a typed architecture graph, builds a working app from it, and verifies on every commit that the code still matches the drawing.
curl -fsSL https://graphlit.co/install.sh | shRun it from anywhere. It asks which folder to turn into a project. Already installed? Run graph setup inside the folder you want.
bash · zsh · sh · dash · fish · Git Bash · no sudo, nothing outside your home directory · read the script first
Free to start · your code stays yours · export any time
No prompt. No template. No prose.
The problem
Everyone ships the ceiling. Nobody ships the floor.
A thousand-line prompt
Boxes and arrows
Code with no link to the ask
Every node → a real file, hashed
Your first hand-edit ends it
A hand-edit becomes a decision
Greenfield only
Import what you already have
The part nobody else has
Three weeks later, is the drawing still true?
Press the button. It runs.
In sync — still hashes to what we built
Changed — hand-edited — accept or rebuild
Gone — the file is no longer there
Unbuilt — drawn, never built
Five nodes claim a file in this repo.
Press Check sync to hash every one of them against the code that is actually on disk right now.
Inside the product
Five surfaces, one graph underneath
Typed nodes and edges, laid out automatically. Select anything to see where it lives.
You do not have to start from a drawing.
Point it at a repository that already exists (a local folder, a git URL, or an upload) and the same graph comes out the other side.
Start where you are
Already have a codebase? Better.
Most tools here need you to start from nothing.
A map that's derived
not drawn in 2024
Rules you can enforce
not a wiki page
Change with a blast radius
scoped to real files
An audit that starts focused
risk by architecture
Under the hood
Mostly not a model, on purpose
Two stages need a model. The other five don't.
Sketch → shapes
deterministic
Shapes → typed nodes
model
Infer what wasn't drawn
model
Graph → ordered work
deterministic
One task, then stop
model
Green or reverted
deterministic
Drawing still true?
deterministic
› 6 shapes, 5 arrows, 6 labels recovered from the board
Watch it
Two minutes, start to finish
Four boxes, three arrows. No prompt.
- 0:00
You draw
- 0:03
It types the graph
- 0:07
It plans the work
- 0:10
It builds one task
- 0:14
The gate runs
- 0:17
It checks the drawing
Everything in the box
Ten capabilities, one underlying idea
Proof, not a promise
We build Graphlit with Graphlit
A tool that keeps architecture honest should survive being pointed at itself. Ours is the first codebase it had to.
Is this another prompt-to-app tool?
No — those stop caring about your code the moment it's generated; this keeps checking it against the drawing.
Do I have to draw? I'm not a designer.
Boxes and arrows, with your finger if you like — or import a repo and never draw anything.
Does it work on my existing codebase?
That's the case it's strongest at.
What happens when the AI writes something wrong?
The task is reverted with the real error — we bound the damage, we don't pretend it can't happen.
Who owns the code?
You do — export it any time, including before you ever pay us.
Do you train on my code?
No, and most of the product never sends it to a model at all.
Can I use my own model provider?
Yes, on any plan — an Anthropic key, an Amazon Bedrock role, OpenRouter, or any endpoint that speaks the OpenAI protocol, with your own address and model names, and yours answers before ours does.
What languages?
Mixed Python and TypeScript today — Next.js, FastAPI, Flask, Prisma, SQLAlchemy, raw SQL.
When can I use it?
Nothing to wait for — create an account and it works straight away. Existing codebases are where it earns its keep fastest.
Point it at something real
Import a repository you already have and see the architecture it derives — and where the code has already drifted from it.
Free to start. Import a repo and the graph is there in minutes.