# unwalled + drawio-mcp: From app clicks to diagrams

Canonical: https://www.kartikaydhar.com/blog/unwalled-drawio-mcp-app-diagrams
Published: 2026-11-11T04:01:35.493+00:00
Updated: 2026-11-11T03:52:41.464601+00:00
Category: Tool Breakdowns

I combined unwalled and drawio-mcp to map a demo app's lookups, records and user flow. Here is how to try it and check the results.

I turned one browser walkthrough of a sample app into an editable map of what it could look up, a diagram of the records it returned and a step-by-step flow to review with an engineer. If you are a PM inheriting an undocumented internal tool or preparing a process review, that is a useful starting point.

unwalled, from KorroAi, watches you use a web app in a browser and saves the requests it makes as reusable tools. jgraph's drawio-mcp lets an AI tool or program draw editable draw.io diagrams. I connected both through MCP, a standard for calling tools, and used a custom test program to build the diagrams.

I supplied no API documentation. The recording gave me enough to draw the app's available lookups, but it left gaps in explaining which page caused each request. I had already checked unwalled's app lookups and draw.io MCP's diagram creation separately. This test put them together.

## What I tested

- I ran unwalled 0.5.1 and the local drawio-mcp diagram server, package version 1.7.0 at commit 639186a.
- I used unwalled's built-in Acme back-office demo. Its customers and orders were fake records in a working sample app, not a real company's data. The demo signed in with its sample account.
- The walkthrough performed six actions after loading the sign-in page: fill the username, fill the password, submit, open an order, filter customers by city and open a customer.
- I used the saved recording to build a lookup map, a customer-and-orders diagram and a flow of the recorded journey.

Recording the demo took about 11 seconds. The separate combined run, including data lookups, checks and browser screenshots of all three diagrams, took about 14 more.

Both repositories use Apache License 2.0, with no software licence fee. This demo needed no paid keys; paid models and connected services can add their own costs.

Image: Illustration: What I tested

The corrected lookup map. Some arrows overlap the boxes and need cleanup.

## What did the lookup map tell me?

An API is how software asks other software for information or an action. The recording found four API lookups: the order list, an individual order, its notes and a customer. It also produced three tools that read information from web pages and a broader tool that could send other requests using the saved demo session.

unwalled calls its saved lookup instructions a recipe. I used that recipe to group the lookups and label each one with its inputs. A PM and engineer could use the map to discuss which reads a workflow depends on and which access is broader than they need.

The map covers what the recording found. It does not reveal the app's internal services or database design. The broader request tool also needs an access review before anyone repeats this with company data.

My first map drew successfully with empty groups because my program misread the response. After I fixed it, all four API lookups and all three web-page tools appeared. A diagram can render fine while missing its contents, so I would check the number of items before reviewing the layout.

Image: Illustration: What did the lookup map tell me?

The first drawing opened, but its groups were empty.

I added a building symbol and an API symbol from Iconify, an online icon library. Both rendered. Several arrows still ran through the groups and looked like one lookup led to the next, even though the diagram connected them back to the app. I would fix those arrows before using the map in a review.

## Could I put the actual records on a diagram?

I fetched the demo's 40 customers and all 200 orders, then grouped the customers across eight cities. Each customer got a card listing that customer's orders, with status, date and total. These were records returned by the running app, not invented labels for a mockup.

I checked each city filter against the full customer list. All eight returned the same customer IDs as the corresponding unfiltered group. I also checked that every customer and order ID appeared in the diagram document.

Image: Illustration: Could I put the actual records on a diagram?

All 200 orders are listed inside the customer cards. Open the image separately to inspect them.

The first saved export held only 100 of the 200 orders. The app returns results in batches, often called pagination. Although I had requested multiple batches, the file held only the first response. I fetched ten batches of 20 explicitly and checked that I had 200 unique order IDs.

Before treating a diagram like a report, compare its record count with the source. I would also split this one by city for a meeting. Listing every order helped me check completeness, but made the full canvas cramped.

## How much of the user journey could I explain?

The final diagram followed sign-in, opening order 1001, filtering customers to Lyon and opening customer 512. Beside the order step, I placed the matching order, notes and customer lookups. Order 1001 was paid, with a total of 948 in data that had no currency field.

One captured list-all-orders request could not be assigned to a specific step. I left it in a separate box. That lets an engineer see the missing connection without mistaking a guessed arrow for recorded behaviour.

The recipe, unwalled's saved lookup instructions, stores reusable paths with one example for each. It does not keep a full request history for every click. The customer lookup example was customer 521, while the later customer page was 512. A fresh lookup could retrieve 512, but could not prove that the original page visit made that request.

Image: Illustration: How much of the user journey could I explain?

The flow keeps the unassigned request separate instead of guessing which page triggered it.

## What this means for product and GTM teams

I would bring the lookup map and flow to an internal process review. The PM can check whether the journey matches the job people do. The engineer can review the inputs, permissions and gaps in the recording. The customer diagram gives them concrete records to compare with the app.

For colleagues explaining the process, I would reduce the record diagram to the customers needed for that discussion. The full version is useful for checking coverage; it is too dense for a slide. I would have someone review the arrows before sharing either diagram.

## How to try it

Start with the demo and an engineer. These steps use the projects' documented local setup routes. I tested a different draw.io server build. The option below opens diagrams directly in your browser.

- Have the engineer install Node.js 22 or newer, then run npx -y unwalled. Its guided setup checks the browser and offers to connect supported apps, including Cursor and Claude Desktop. Approve only the app you want to use.
- Choose the demo in that setup, or run npx -y unwalled demo. Keep it running while you try the lookups.
- Add the local draw.io package to the same client. Its README supports Cursor and Claude Desktop: use npx as the command and ["-y", "@drawio/mcp"] as its arguments. The client starts that package. Enable both connections and check that their tools appear before proceeding.
- First ask the README's demo question: "Using acme, what is the status and total of order 1042? And which customers are in Lyon?" Compare the answer with the sample app.
- Then ask for a map of the discovered lookups, followed by a small customer diagram and a flow of the recorded pages. Ask it to show counts and leave uncertain request-to-page links unassigned. My tested combined flow needed a custom script to read the results, group records and describe the diagrams; this prompt is a suggested trial, not a verified replacement for that script.

Skip this if no engineer can review the result. For the local option above, draw.io's README says the diagram stays in the link and is not sent to its servers, though the editor and remote icons still load from the internet. If your data rules forbid that, skip this route or ask your engineer about local and self-hosted options.

## Frequently asked questions

### Can I use this for an undocumented internal app?

That is a possible use for the browser-recording route. My test used the built-in demo without supplying API docs; review permissions and try one known workflow before using your own app.

### Will connecting both tools automatically produce these diagrams?

My test needed a custom program to handle the results and build the diagram descriptions. The setup above connects the tools, but does not reproduce that program.

### Does the saved recording explain every click?

No. It keeps lookup paths with examples, so some requests cannot be assigned to a particular page. I left the list-all-orders request unassigned.

### What should I check before sharing a diagram?

Check record counts, arrow relationships and private data. My empty first map and half-sized first export both showed why opening a diagram is not enough.

## Sources

- unwalled README: browser recording, demo, setup, supported clients and Node requirement. Checked 11 Nov 2026 (IST).
- drawio-mcp README at the tested commit: editable diagrams, local and hosted options, and data handling. Checked 11 Nov 2026 (IST).
- Local draw.io tool-server README: package setup for Cursor and Claude Desktop. Checked 11 Nov 2026 (IST).
- unwalled licence and drawio-mcp licence: Apache License 2.0.
