Diagram as code is excellent once you already know the answer
Mermaid, PlantUML, D2 and Graphviz are free, version controlled, and render straight out of a repository. For a diagram that belongs next to code, nothing else comes close. They share one assumption: you can already describe the process precisely enough to type it.
During discovery you cannot. You are learning the process in fragments from someone who has never had to explain it in order before, and every second spent on syntax is a second not spent listening. That gap is what Wavelength is for.
The demo needs no account. The free plan needs no card.
The actual difference
| Diagram as code | Wavelength | |
|---|---|---|
| Input | Syntax you type | Speech, or a pasted transcript |
| What it assumes | You know the flow well enough to write it down | You are hearing it for the first time |
| Where it lives | A file in your repository | A hosted flow on your account |
| Version control | Native. Diffs in a pull request | Version history and restore on Pro |
| Layout | The rendering engine decides | Auto-laid out, then draggable by hand |
| Shows its uncertainty | Not applicable, you typed it | Yes, inferred links render dashed amber until confirmed |
| Usable live in a meeting | Only if you can type it faster than they talk | Yes, that is the design point |
| Cost | Free and open source | Free tier, then $9.99 per user per month |
Mermaid, PlantUML, D2 and Graphviz are free and open source. Commercial hosted editors built on top of them are priced separately. Cost is not the deciding factor here.
Where diagram as code is still the right call
Keep writing diagrams as code when the diagram is documentation that should rot at the same rate as the code, which is to say when it should be reviewed in the same pull request that changes the behaviour it describes. Nothing hosted gives you that.
Keep it when the diagram is generated. If a script can emit the syntax from a schema, a config, or a service registry, you have a diagram that is correct by construction, and a person narrating it is a downgrade.
Keep it when the audience is developers and the rendering target is GitHub, GitLab, Notion or Obsidian, all of which render Mermaid with nothing to install.
And keep it when you want an artifact with no vendor, no seat, and no account. That is a real property and it is worth something.
Where typing costs you the meeting
Someone says this on a discovery call:
"Leads come in through HubSpot, marketing qualifies them, and if they look good they route to an SDR. Meanwhile everything syncs nightly into Snowflake where dbt cleans it before it hits Tableau."
To turn that into Mermaid you have to make four decisions before you type a character: which diagram type, which node ids, which of those things are decisions rather than steps, and whether the nightly sync is a branch or a separate graph. Every one of those decisions is thinking you are doing instead of listening.
flowchart TD
HubSpot[HubSpot] --> Mktg[Marketing]
Mktg --> Q{Qualified?}
Q -->|Yes| SDR[Sales development rep]
Q -->|No| Nurture[Nurture campaign]
HubSpot --> Snow[(Snowflake)]
Snow --> dbt[dbt] --> Tableau[Tableau]
That is not hard to write. It is impossible to write at conversational speed while also running the call, and it does not record that you inferred the HubSpot to Snowflake edge rather than hearing it stated.
Wavelength takes the sentence and produces the same structure without the four decisions. Systems, actors, a real decision branch, and the nightly sync as a parallel path. Then you correct it by voice, by prompt, by right-click, or by hand.
The dashed amber lines
Code has no notion of confidence. Every edge you type reads as fact. Wavelength draws inferred connections dashed and amber, so before you send the diagram to a customer you know exactly which two or three links to confirm out loud.
Using both
The honest workflow is sequential. Capture live in Wavelength while the process is being described, correct it in the room, and get agreement. Then, if the diagram needs to become a durable engineering artifact, write it as Mermaid in the repo where it will be reviewed and updated.
That second step is manual today. Wavelength exports PNG, a shareable link, and an email recap on Pro. It does not emit Mermaid or PlantUML syntax. If code export would change your decision, tell us, because that is the kind of thing that gets built when enough people ask.
Common questions
Could I not just ask an LLM to write the Mermaid for me?
You can, and for a tidy paragraph it works. Two things are different here. The diagram appears while the conversation is happening rather than after it, so the other person can correct it in the room. And inferred edges are marked as inferred rather than silently becoming fact, which is the failure mode of pasting a transcript into a chat window and trusting the output.
Does Wavelength do sequence diagrams or class diagrams?
No. Process and data flow only, using six shapes. For UML, keep PlantUML or a general diagramming suite. See the Visio comparison.
What if the meeting already happened?
Paste the transcript and you get the same diagram. See transcript to flowchart.
Can I edit the layout?
Yes. Nodes auto-arrange and then drag freely, and connections can be reversed or reshaped by clicking them. This is one of the more common complaints about code-rendered diagrams, where the engine's layout is what you get.
Try it before you decide
Describe a process you already know well enough to write as Mermaid, and compare what comes back. That is the fairest test of whether the capture step is worth anything to you.