01About
The story behind the name
This app is dedicated to Larry Jeffus.
Larry is an engineer with a deeply inspiring story. He began life with severe dyslexia, and although he could not read, he went on to write more than 40 books and to teach hundreds of thousands of students around the world.
He loves to make all his engineering designs with diagramming applications. That is why his niece, Janet Jeffus, decided to make this app and dedicate it to his inspiring story.
02The project
Why Larrygram exists
Diagrams outlive the tools that made them. Larrygram is a clean-room, open-source diagramming app that reads and writes Visio .vsdx files, so drawings stay open to anyone who needs them.
It is built only from public specifications, Microsoft Learn reference pages, published control identifiers, public support articles and observed behavior. It contains no Microsoft code, binaries, icons, fonts, stencils or templates.
Larrygram is also designed to work with AI agents. A planned MCP server will let agents open drawings, run the app’s commands and read the results, through the same command registry the app uses. Until it ships, agents and scripts can use the command-line tool. More about the MCP server.
03How it is built
Written in Rust, by AI agents, in the open
Larrygram is written in Rust and built by AI coding agents, including Claude Code and pi, working under the written rules in AGENTS.md. Each agent works on its own branch. A master agent reviews every change and merges it.
Claude Opus 5.5 acts as the overseer agent. It hands out jobs by complexity: MiMo-V2.6-Pro takes large and medium jobs, and DeepSeek V4.1 Flash and Kimi K2.6 take medium and small ones. Together the three models have used 3.75 billion tokens, and the entire project has cost less than $100 in token costs on OpenRouter.
Every merge passes a quality gate: formatting, lints, documentation, license checks, the test suite, coverage thresholds, a WebAssembly build and fuzz targets. The last full gate ran 2,835 tests.
Pull requests are also reviewed by hand by senior software engineers at Maitri Group, LLC.
Built by AI agents for under $100.
- Tokens across three models
- 3.75 billion
- Token costs on OpenRouter
- Under $100
- Overseer agent
- Claude Opus 5.5
Share of tokens by model
- MiMo-V2.6-Pro80.5%
- DeepSeek V4.1 Flash15.2%
- Kimi K2.64.3%
Jobs go to models by complexity. MiMo-V2.6-Pro takes large and medium jobs; DeepSeek V4.1 Flash and Kimi K2.6 take medium and small ones.
Figures as of October 23, 2026.
04Credits
Credits and acknowledgments
App icon and illustrations by Janet Jeffus, made in VectorCraft.
- libvisioReference for the planned read-only .vsd import (MPL-2.0).
- egui and eframeThe user interface toolkit for the desktop and web apps.
- tiny-skiaPNG rendering.
- Carlito and SelawikOpen fonts bundled for drawing text (SIL OFL 1.1).
- Tabler IconsThe app’s command icons (MIT).
- Bricolage Grotesque, Geist and Noto Sans ThaiThe fonts of this website (SIL OFL 1.1).
Contributors
The contributor list appears here once the repository is public.