Avgidea Adds Local LLM Coding to Greenative Studio for Mac
Coding agents such as Claude Code can now run against
Local Models on Apple Silicon
When we launched Greenative Studio earlier this year, one question came up again and again in conversations with customers: "This is great for chat — but can it work with our coding tools?" It was a fair question. For many of the businesses we talk to, source code is the single most sensitive asset they own, and yet agentic coding tools almost universally assume you're willing to stream that code to a cloud API. We didn't think that trade-off should be mandatory.
Today's Coding release is our answer. Below is the full announcement we published this morning, as covered by News.com.au and other outlets.
Avgidea, Inc. announced a major update to Greenative Studio, its private AI workbench for Mac. The new Coding release adds an Anthropic-compatible API gateway that lets agentic coding tools — including Claude Code — run against large language models executing locally on the user's Mac, and introduces support for NVIDIA's newly released Nemotron 3.5 Lightning model.
For many businesses, source code is among their most sensitive intellectual property. As agentic coding tools become part of everyday development work, more teams in Australia and New Zealand are weighing the productivity gains against a simple concern: every prompt may carry proprietary code to a cloud service. The Coding update gives those teams another option — running the same agentic workflows against a model that never leaves their machine.
With the update, developers can point Anthropic-compatible coding tools at a local endpoint served by Greenative Studio. In internal testing, Claude Code connected to a locally running model was able to explore a multi-thousand-line codebase — reading files, running shell commands, recovering from errors, and producing an accurate architectural summary — with no code or prompts sent to any cloud AI service.
"Source code deserves the same protection as any confidential business document," said Teppei Yagihashi, founder of Avgidea. "Agentic coding is transforming how software gets built, but it shouldn't require sending your repository to someone else's servers. With this update, the agent's brain runs on your own Mac — your code, your prompts, and your model all stay local."
Key additions in the Coding release include:
An Anthropic-compatible API gateway for agentic coding tools, compatible with Claude Code and other Anthropic-compatible clients
Support for NVIDIA's Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model designed for agentic workloads such as tool calling, code review, and task execution
An expanded model catalogue, adding Google's Gemma 4 26B (A4B) and Qwen-family models
A real-time token metrics dashboard showing generation speed and cumulative token usage for local coding sessions
One-click model start-up — no Python, Homebrew, or command-line configuration required
Greenative Studio continues to focus on curated open-weight models released under permissive licences (e.g. Apache 2.0, MIT, OpenMDW), helping businesses evaluate local AI workflows with clearer commercial-use expectations.
The largest models in the catalogue require an Apple Silicon Mac with at least 32 GB of unified memory. In Avgidea's testing on an M4 Max Mac with 128 GB of unified memory (Q8 build), Nemotron 3.5 Lightning sustained approximately 70–80 generated tokens per second during agentic coding sessions. Lighter models, including the free Phi-4-mini plan, remain available for Macs with less memory.
The Coding release is included in the Greenative Studio Pro plan at US$18 per month (approx. AU$28) or US$180 per year, with fixed pricing and no per-token usage charges for local workflows. A free plan with Phi-4-mini is available, and new users can start without a credit card.
Product landing page: https://www.avgidea.io/en/private-ai
Free download: https://www.avgidea.io/en/downloads
Comments