About coccinella-labs

[bniladridas]

coccinella-labs builds small developer tools and the systems behind them. Published work is public; training sources for the Hub models are kept private.

Work groups into five systems.

Agent infrastructure

Harnesses, runtimes and bots that plan work, execute it, and ship results.

• harper - Rust agent runtime with a TUI, an HTTP API and a sandbox
• agent-sdk - reusable agent runtime extracted from Harper
• agentware - C++20 AI coding agent for the terminal, reading and editing code through MCP filesystem and git servers
• harperbot - GitHub App bot for automated code review
• swe-agent - software engineering agent for this organization

GPU and machine learning compute

Apple Silicon kernels, Metal runtimes and inference for local and serverless machine learning.

• core - Metal GPU compute runtime focused on memory, synchronization and data movement
• kernels - Metal compute kernels in Swift and Metal, with a runnable CUDA translation demo
• gpucomm-fs - content-addressed store for GPU artifacts, with verified retrieval and automatic deduplication
• gpucomm-bot - GPU-aware GitHub App and CI automation
• bitinfer - Hugging Face encoder inference for Apple Silicon, halving memory with float16 weights and slower than plain transformers on CPU
• ml - MPI coordination scaffold for distributed training in C++, with a REST monitoring dashboard. The learning itself is not implemented.
• mlapi - FastAPI service for model inference
• rl - CMA-ES reinforcement learning for CartPole-v1 with a linear policy

Where to look next

The catalog at coccinella-labs.github.io lists every code project with its language, type, license and install command.

Organization-wide defaults, community health files and the git-gui package live in .github.

Note

This post used to describe an organization named Harper. That name is
historical. The organization is coccinella-labs, and the repositories above
are the current catalog.