By Tuna, the maker of FlowDiagram Every engineering team I have worked with has the same graveyard: a Confluence page with an architecture diagram that was last true two re-orgs ago. Somebody drew it in a heavyweight tool, exported a PNG, pasted it in, and the source file now lives on a laptop that was returned to IT in 2023.
I'm a data architect. I spend a lot of my week drawing three kinds of pictures: integration landscapes (which system talks to which, over what), data pipelines (source → ingest → transform → serve), and ER models. I wanted one tool that did all three well, that didn't make anyone sign up to look at a diagram, and that an AI assistant could write for me without screenshots or plug-ins.
I couldn't find it, so I built it. It's called FlowDiagram, it lives at flowdiagram.net, and it is free: no tiers, no trial, no account.
Here is what makes it different, and why I think it is the right shape of diagram tool for 2026.
Click Share (Ctrl+L) and you get a URL. That URL doesn't point at a file on my server. The diagram itself is compressed into the part of the URL after the #, which browsers never send to any server.
So:
Don't take my word for it. Each of these opens a complete diagram:
Opening a link never replaces your work; it adds the diagram as a new page.
Most "AI diagram" tools put a model behind a paywall and give you a prompt box. I went the other way.
FlowDiagram publishes a single Markdown skill file, flowdiagram.net/flowdiagram.md, that teaches any coding assistant the document format. Drop it into Claude Code, Codex, Cursor, GitHub Copilot, Zed or Windsurf, then just ask:
"Draw our order pipeline: Shopify and Stripe into Kafka, dbt in Snowflake, Power BI on top."
The assistant writes a .flow.json file and a one-click link. Because the output is text, there is no image-model pass and no API key on my side. And because you keep the .flow.json, the assistant can revise the diagram later instead of regenerating it from scratch.
Why a skill file and not an MCP server? A server has to be installed and kept running, and its tool definitions ride along on every request. A skill file is one Markdown file that is only read when you actually ask for a diagram. It can't go down.
Architecture diagrams are about which systems, so generic boxes labelled "DB" don't cut it. FlowDiagram ships with 1,160 vendor logos, including every AWS (303), Azure (405) and Google Cloud (175) service icon, plus SAP, Salesforce, Snowflake, Databricks, Kafka, dbt, Airflow, Power BI, OpenAI, Anthropic and hundreds more. Press B for the library or Ctrl+K and type a name.
Have an internal system with its own logo? + Logo adds it to your library for good.
The entity mode (M) gives you proper tables: typed columns, primary and foreign keys, column comments, and crow's-foot cardinality that attaches to the actual column rows.
And it isn't a dead-end drawing:
CREATE TABLE, DBML, Mermaid erDiagram, a CSV header, or a live Paste DDL in, get a diagram out, edit it, ship DDL back. That loop is the reason I started this project.
Turn on Flow for any connector and packets travel along it in the direction the data moves; hollow packets show return traffic. Click any system and FlowDiagram traces everything downstream of it. Save views (bookmarked zoom positions) and hit Present: the live canvas becomes your slide deck, so animations keep running and nothing goes blurry when you zoom.
This is the feature that makes architecture reviews faster. "What breaks if this queue stops?" becomes one click.
You don't have to redraw anything:
Export goes to PNG, SVG, PDF, Mermaid, DBML, SQL and .flow.json.
There is no account and no sync service, so there is nothing to breach. After the first visit the whole app is cached and works offline (everything except the live Snowflake connection). Install it from the address bar and it runs as a desktop or mobile app. Analytics are cookie-free page counts; your diagram content is never touched.
Keep in mind the trade-off that comes with this design: there's no real-time co-editing or team workspace. You collaborate by sending links and committing .flow.json files to Git, where they diff like any other text file. For architecture docs that live next to code, I think that's the right trade.
Or let your AI do it: give your assistant flowdiagram.md and describe the system in plain English.
I build FlowDiagram on my own, so feedback shapes it directly. If something's missing, the support address is in the docs. Tell me what you'd draw first.
Originally published on Medium.