MMMatthew Mercado
Resume
All case files

Case 02AI SUPPORT AGENT

Wicket

A help and feedback chat that drops into any web app with one script tag. An AI assistant answers from the product's docs, and every conversation lands in the team's Discord, where a reply goes straight back to the person in the app. It's free and open source under the MIT license.

Role
Creator & Lead Software Engineer
Timeline
2026 – Present
Status
Live and growing
Links
Visit the live product View the code on GitHub
1
Script tag to add it to an app
27
Real-world questions in its model test set
24 KB
Widget size, gzipped
Wicket interface

01 Why build it

Support for a small software company usually means a help desk priced per seat, a chatbot that can't see the product, and bug reports that disappear once someone replies. I wanted customers to get an answer in seconds from our own docs, reach a person in one tap, and hear back when the thing they reported is fixed, all from tools the team already has open.

02 How it works

One message in, the right answer out.

  1. Answers from the docs, with receipts

    Wicket indexes a Markdown docs repo or any docs site's sitemap into SQLite full-text search and links the exact section it used. Through app tools, the assistant can also read the signed-in person's own data from the product's backend, checked against their token on every call, and it shows the numbers behind each answer.

  2. Discord is the inbox

    Each conversation becomes a private forum post in the team's Discord. A reply there, or in the web dashboard, reaches the person in the app live. When someone asks for a person, gets frustrated, or the docs don't cover it, the assistant hands off and pings the team.

  3. Everyone hears back

    The assistant files bugs and ideas in Linear or GitHub Issues, adds repeat reports to the ticket that already exists, and tells every reporter when it's fixed or planned, in the chat or by email if they've gone.

  4. Cheap and safe to run

    It's one Node process with SQLite. The widget renders in a shadow DOM so it can't break the host app, identity comes from short-lived signed tokens, and every person has a monthly AI budget. Models are compared with an eval harness I wrote that scores replies without an AI judge.

03 Launch trailer

One report, start to finish.

The launch trailer One bug report, told in type: answered from the docs, filed, counted, shipped, and sent back to everyone who reported it.

04 The site

Logs the bug
Logs the bug A scripted demo on the site: the assistant files the bug with the page, version and console error, then tells the reporter when the fix ships.
Nobody's report disappears
Nobody's report disappears Repeat reports join the ticket that already exists, and everyone who reported it hears back.

05 Built with

  • TypeScript
  • Node.js
  • Hono
  • SQLite
  • Drizzle ORM
  • Vercel AI SDK
  • Claude
  • discord.js
  • React
  • Railway
Next case fileEnterprise AI Agent Tooling