Metalware

Metalware is a privacy-first AI programming copilot. All processing happens 100% locally, meaning no sensitive data leaves your computer. Metalware is an excellent choice for industries that prohibit cloud-based AI services, such as aerospace and defense.

Approach

We're starting with copilot features for firmware, tools that we wish we had at SpaceX (see Our Story section below!). Additionally, here's what differentiates our product:

• We use custom shaders to leverage Apple silicon and maximize GPU performance.
• We train small, specialized LLMs for individual use-cases, which outperform current state-of-the-art large models on standard benchmarks.

Features

• Automated documentation / docstrings
• Q&A search for code + datasheets
• Context-aware code generation:
  • Multi-line autocomplete
  • Unit tests
  • HALs
  • Boilerplate
• Debug registers and real-time system memory
• Static analysis

Metalware for Individuals

• Runs 100% locally, offline
• Native macOS app + editor plugin
• Optimized for Apple Silicon (M1 or M2)

Metalware for Enterprise

• Support for larger codebases
• Self-managed VPC or server
• More powerful Large Language Models (LLMs) + fine-tuning
• SOC 2 compliance
• Organization-wide license + policy management
• Corporate proxy support
• SAML-based Single Sign-On

Benchmarks

Coming soon...

Language support

• C, C++, Python
• (coming soon!) Rust, JavaScript, Java, Kotlin, PHP, Ruby

IDE support

• VS Code
• (coming soon!) Visual Studio, Vim, Neovim, JetBrains, Emacs

Data and privacy

• What data has Metalware been trained on?
  • Metalware has been trained on permissively licensed code available on the internet.
• Will my code be shared with anyone?
  • No. Your code does not leave your computer. This is verifiable with a packet capture.
  • *Many alternative products claim data privacy, but still send contextual "snippets" of your code to cloud-hosted LLMs.
• What data does Metalware collect?
  • Number of suggestions generated + accepted.
  • System OS version.
  • Average inference latency.
• How is the data in Metalware used?
  • The data is used to track how well the suggestions are working for our users across programming languages and operating systems.

Trusted by engineers at leading companies :)

Our Story

Hello! We're Andrew and Ryan. We met at SpaceX, where we helped start the Starlink factory and wrote firmware that runs on more than one million antennas globally. In June 2023, we were accepted into Y Combinator and jumped into startups.

We wanted to build tools that we wish we had at SpaceX, so we started working on an idea to automate manual, error-prone firmware tasks involving dense technical data sheets.

After talking to dozens of engineers, we quickly realized many companies today can't use AI tools due to data privacy concerns. We knew we needed to bring privacy-first AI to enterprises to unlock vast end-user value.

Long-term, we believe AI models will be customized for every use case. Open-source models are improving at a blistering pace, and we're excited to help accelerate AI at the edge.

We're early in our journey and would love to discuss your ideas and feedback! Feel free to say hello at ryan@metalware.cc