Skip to content
Karl Foster

Builds · apps, agents, tools

I still build.

Here are some fun things I’ve built in my spare time.

01 · My own builds

01 · MCP server

WhatsApp MCP

A private tool that lets my AI assistant search and read my own WhatsApp, voice notes included.

Status
Private build

It keeps a copy of my chats on my Mac and lets the assistant search and read them. Voice notes are transcribed on the Mac too, so nothing leaves it. I built it because my working day runs on WhatsApp. It stays private, because it holds other people’s messages.

Stack: TypeScript · Node 24 · SQLite · launchd · Parakeet TDT · MCP

02 · macOS app

Lovox

Push-to-talk dictation for the Mac. It runs on the Apple Neural Engine, so your voice stays on your Mac.

Status
Live

Hold a key, talk, let go, and the text appears at your cursor. I built all of it: the app, the speech pipeline, the auto-updates, the licensing and the website. It shipped 28 signed releases in its first three months.

Stack: Swift · SwiftUI · Core ML · Parakeet TDT · llama.cpp · Sparkle

03 · Open source · MCP server

claude-gpt

An open-source tool that lets Claude hand heavy work to OpenAI models and bring the answer back.

Status
Published

Claude passes the job and its files to an OpenAI model and brings the answer back into the same chat. Jobs run in the background, and each one is costed against a budget.

Stack: Python 3.12 · MCP SDK · OpenAI Responses API · SQLite · aiohttp · GitHub Actions

From the README: “Built by Karl Foster with Claude.”

04 · Open source · Notebook

Gym Tinder

An algorithm that pairs gym members with training partners.

Status
Published

Members answer a few questions about their goals, their experience and when they train, and the algorithm pairs them up. A gym that introduces two members who train at the same hour gives them one more reason to keep coming.

Stack: Python · Jupyter · pandas · Euclidean matching · Circular time encoding · Synthetic data

05 · Fitness · AI agents

Neural PT

Nine AI agents that write a personalised six-week training programme.

Status
Lab build

I loved being a personal trainer, but I hated writing programmes. Here you fill in a questionnaire, nine AI agents plan and check the programme, and a finished six-week plan arrives by email in about nine minutes. When I wrote it up, each one cost around 70 cents to produce.

Stack: Python 3.13 · FastAPI · Google ADK · Gemini via LiteLLM · Supabase · Docker

06 · Open source · Tools

Claude Board and the macOS utilities

Small tools I built for my own working day.

Status
Two public, four private

I built these because I needed them, and I use them every day.

Stack: Swift · SwiftUI · PDFKit · CoreAudio · Next.js · MCP SDK

  • Claude BoardA Kanban board that lives inside Claude Desktop over MCP, so I can move cards by asking.GitHub
  • WorkflowsRuns apps, commands and text snippets from one hotkey.
  • InkwellA clean PDF editor that fills forms and stamps signatures.Private
  • TaskAppKeeps my tasks in a plain-text file.Private
  • QuitAll · AirPodsShortcutQuitAll quits every open app. AirPodsShortcut connects my AirPods from a hotkey.Private

Want to work with me?

If you’re weighing up something like this, it starts with a short call.