3-tier web app on AWS
Classic load-balanced web + app tier over a cache and a managed Postgres, in a VPC.
MCP-native · works inside Claude Code, Cursor & WindsurfDescribe a system.
Describe a system.
Get a diagram you can actually edit.
Flowgraf turns a sentence into a clean, editable cloud-architecture diagram. Change the design in plain English — it re-wires and re-lays-out itself, no dragging boxes around.
Free to start · AWS & generic icons · Export SVG · PNG · Mermaid
flowgraf / canvas
↳add a Redis cache between the API and the DB
How it works
Two front doors. One diagram engine.
Whether an AI agent or a human authors the change, it becomes the same validated operation against one source-of-truth diagram — then a deterministic layout engine draws it. No model ever places a box by hand.
For AI agents
Diagram from inside your agent
Add Flowgraf to Claude Code, Cursor, or Windsurf. Ask for an architecture diagram — your agent calls the tool and hands back a link to a live, editable canvas.
terminal
$ claude mcp add --transport http flowgraf https://flowgraf.in/api/mcpCursor & generic stdio: npx flowgraf-mcp · full install docs →
For humansOpen a blank canvas →
Or just describe it in the browser
Open a blank canvas and type what you’re building. Every follow-up — “put a queue in front of the workers” — re-wires it instantly. Drag to fine-tune; pinned nodes stay put.
a 3-tier web app on AWS with an ALB
Added ALB → 2× EC2 → RDS, grouped in a VPC ✓
add CloudFront in front
Reviews its own work
One click, and it critiques the diagram — visually and logically — before you ship it.
Hit Review and a vision model looks at the actual picture and the underlying graph: overlapping edges, cramped spacing, a confusing flow direction — and the logic too, like a service talking straight to a database or a missing cache. It asks about anything ambiguous, then applies validated one-click fixes.
It never nudges a box by hand — every fix is a semantic operation the layout engine re-draws. The model critiques the picture; it never moves a pixel.
✦ Review2 visual · 1 logical · 1 question
highmissing_link
The web app talks straight to the database.
Insert a cache between the API and the DB.
mediumspacing
Two nodes overlap near the gateway.
Unpin & re-layout, or increase spacing.
Apply all safe fixes
Templates
Start from a real diagram, not a blank canvas.
Pick a pattern, click once, and it’s yours to edit — by hand, in chat, or from your agent. Every one is a live diagram, laid out by the same engine that builds yours.
Event-driven microservices
An event bus fans requests out to independent services with a dead-letter queue.
RAG pipeline
Retrieval-augmented generation: embed a query, search a vector store, ground the LLM.
Kubernetes deployment
Ingress → Service → Deployment with a StatefulSet database, inside a cluster.
Serverless API
API Gateway with JWT auth, a Lambda handler, DynamoDB, and an uploads bucket.
Data lakehouse
Streaming and batch sources land in a data lake, curated into BigQuery for BI and ML.
CI/CD pipeline
A build → test → deploy flowchart with a decision gate and an alert branch.
Pricing
Free to start. Pro when you live in it.
Launching free. Pro is opening soon — join the early-access list and lock in the intro price.
Free
$0/ forever
For trying it out and small projects.
- 25 AI generations / month (5/day before you sign up)
- 3 saved diagrams, kept forever
- Unlimited MCP diagrams from your agent
- Export SVG, PNG & Mermaid (with a small badge)
- Share by link · drag-to-edit canvas
Early access
Pro
$12/ monthintro $9
For people who diagram every week.
- 1,000 AI generations / month
- Unlimited saved diagrams
- Clean exports — no badge
- Private diagrams
- Everything in Free
No charge now — we’ll email you when it opens.
FAQ
Questions, answered
Do I need an account to start?
No. Start on a blank canvas anonymously — describe a system and edit it right away. When you want to keep it, sign in and claim it in one click. Diagrams from your AI agent come with a claim link built in.
Which AI tools does the MCP server work with?
Any MCP client — Claude Code, Cursor, and Windsurf today. Add it once (a remote HTTP endpoint or the npx stdio proxy) and your agent can create and edit diagrams and hand you back a live link.
Does it burn my tokens when my agent calls it?
No. On the MCP path your agent authors the diagram operations itself — Flowgraf just validates, lays out, and renders. There's no model call on our side, so it's free no matter how often it's hit.
What clouds and icons are supported?
AWS services and a set of generic architecture primitives (service, database, cache, queue, gateway, function, and more) ship today. GCP, Azure, and Kubernetes icon sets are on the way.
What is the AI Design Review?
One click and Flowgraf reviews its own work — visually and logically — before you ship it. A vision model reads the rendered picture and the underlying graph, flags overlaps, cramped spacing, confusing flow, missing links and architectural smells, asks about anything ambiguous, and applies validated one-click fixes. It never moves a box by hand: every fix is a semantic operation the layout engine re-draws. It's a signed-in feature (it runs our vision model), and a review uses 3 of your monthly AI generations.
Can I still edit by hand?
Absolutely — it's a real canvas. Drag nodes, draw edges, rename, delete. Anything you position by hand is pinned and survives the automatic re-layout when the diagram changes.
How is this different from draw.io, Lucidchart, or Mermaid?
Those are editors you drive by hand (or one-shot text-to-picture). Flowgraf keeps a semantic model of your system, so a change like “add a cache in front of the DB” re-wires the graph and re-lays it out cleanly — the editing is the point, not just the first draft.
Is my data private?
Diagrams are shareable by link by default (only people with the URL can see them). Fully private diagrams are a Pro feature.