MinuteMaker

A hot-folder automation tool that converts raw meeting notes into formatted LaTeX/PDF documents. Drop a .txt file into the input folder and get a polished PDF out.

MinuteMaker supports both cloud (Anthropic Claude) and local (Ollama, llama-cpp, vLLM, or any OpenAI-compatible endpoint) LLM backends, and uses a configurable template system so any organisation can produce minutes in their own format.

How it works

1. A file watcher monitors the input/ folder for .txt files
2. When a file arrives, MinuteMaker sends it through the processing route for the configured provider
3. anthropic generates complete LaTeX directly, while openai_compatible returns structured JSON (sections, items, metadata)
4. For openai_compatible, MinuteMaker renders the JSON into LaTeX using the selected template
5. pdflatex compiles the .tex into a PDF
6. Output files appear in output/, one self-contained folder per meeting

Quick start

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Then run the setup wizard:

python minutemaker.py --init

This walks you through choosing a template, LLM provider, and generates your config.yaml.

Cloud mode (Anthropic)

export ANTHROPIC_API_KEY=your-key-here
python minutemaker.py

Local mode (Ollama)

# Install Ollama first, then configure MinuteMaker:
python minutemaker.py --provider openai_compatible --base-url http://localhost:11434/v1 --model gemma2:27b

Initial Ollama setup is still required. MinuteMaker assumes:

• Ollama is already installed from ollama.ai
• the ollama command is available on your PATH
• the Ollama API is reachable at your configured base_url

After that:

• if managed_server: true is set in config.yaml, MinuteMaker will start and stop Ollama automatically using server_command (by default ollama serve)
• if managed_server: false, you must start Ollama yourself before running MinuteMaker, for example with ollama serve
• the configured model is pulled automatically on first use, so no manual ollama pull is needed

To switch models, change the model line in config.yaml. A menu of recommended models with RAM requirements is included in the config file.

If managed_server: true is set in config.yaml, MinuteMaker will start and stop the Ollama server automatically, freeing GPU/RAM when not processing.

Verify local setup before cron

Before you enable cron or any other background startup automation, run:

python verify_local_llm_setup.py

This script is designed for non-experts and runs the local setup checks in order:

• confirms Python dependencies and pdflatex are installed
• checks that config.yaml is set to local mode and points at a model/base URL
• runs a small live prompt against your configured local model
• starts a temporary watcher, drops in the bundled sample notes file, and confirms that MinuteMaker produces both .tex and .pdf output

The first run may take a while if Ollama still needs to download the model. If the script finishes with all checks passing, your local setup is in good shape and you can enable cron with much more confidence.

At the end, the verifier prints the absolute paths it found for Python, pdflatex, and the Ollama binary. Those are the paths you should prefer in cron jobs and in server_command, because cron often has a much smaller PATH than your normal terminal.

Local LLM and data privacy

When using a local LLM provider, all processing happens entirely on your machine. No meeting notes, generated content, or metadata are sent to any external service. This makes the local mode suitable for handling sensitive or confidential meeting content and compatible with GDPR and institutional data protection requirements.

MinuteMaker does not store or log the content of your notes beyond the generated output bundles in output/. The local LLM server is started on-demand and shut down after processing, so no model remains resident in memory between jobs.

Usage

# Watcher mode — monitors input/ for .txt files
python minutemaker.py

# Process a single file
python minutemaker.py --once path/to/notes.txt

# List available templates
python minutemaker.py --list-templates

# Use a specific template
python minutemaker.py --template liverpool

Requirements

• Python 3.10+
• pdflatex on PATH (TeX Live, MacTeX, or MiKTeX)
• An Anthropic API key (cloud mode) or a local LLM server (local mode)
• For Ollama local mode: Ollama installed locally, with the ollama CLI available if you want MinuteMaker to manage the server for you

Templates

MinuteMaker ships with two templates:

• default — generic meeting minutes (Welcome, Updates, Discussion, Action Items, AOB, Next Meeting)
• liverpool — University of Liverpool Nuclear Physics Group format

To create your own template, see CREATING_TEMPLATES.md.

Configuration

All settings live in config.yaml:

template: default                       # template directory name
provider: anthropic                     # anthropic=direct LaTeX, openai_compatible=JSON pipeline
model: claude-sonnet-4-20250514
max_tokens: 8192
input_dir: ./input
output_dir: ./output
log_level: INFO

# For local LLMs:
# base_url: http://localhost:11434/v1
# managed_server: true
# server_command: "ollama serve"
# server_startup_timeout: 120

provider now also determines the processing mode: anthropic uses the direct single-pass LaTeX route, and openai_compatible uses the multi-pass JSON route.

For Ollama specifically:

• managed_server: true means MinuteMaker will run ollama serve for you
• managed_server: false means you must already have Ollama running at base_url
• model download is automatic once the Ollama server is reachable
• if you plan to run MinuteMaker from cron, prefer an absolute server_command, for example "/absolute/path/to/ollama serve"

CLI arguments override config file values. See python minutemaker.py --help for all options.

Each successful run creates a self-contained bundle in output/, for example:

output/
  minutes_2026_04_15/
    minutes_2026_04_15.tex
    minutes_2026_04_15.pdf
    header/
      title_code.tex
      liverpool-logo.png

That folder can be copied to another machine and recompiled directly with pdflatex.

Project structure

verify_local_llm_setup.py  # local LLM preflight + watcher smoke test
minutemaker/          # Python package
  cli.py              # CLI entry point, --init wizard
  config.py           # Configuration loading
  pipeline.py         # Core processing pipeline (Anthropic direct LaTeX, local models via JSON)
  providers.py        # LLM provider abstraction
  server.py           # Local LLM server lifecycle + auto model pull
  renderer.py         # JSON → LaTeX rendering
  schema.py           # JSON schema + validation
templates/
  default/            # Generic template
  liverpool/          # Liverpool Nuclear Physics template