Localcode, open source agentic coding on your Mac with one command LocalCode, an open-source agentic coding tool from developer mjwsolo, launched as a one-command install (`uv pip install localcode`) that runs local models on Apple Silicon Macs running macOS 13 or newer with at least 16 GB of unified memory and Python 3.10 or newer. LocalCode offers 12 models and 169 quantization options downloaded from Hugging Face, and adds optimizations over the stock model server including vendor speculative drafting (about 2x faster replies on Gemma 4, Qwen 3.6, Qwen 3.8 and Muse Glimmer), copy drafting (3.5 to 4x faster on file rewrites), an 8-bit KV cache by default, and fixes that move Inkling 2's experts onto the GPU, which the project says were silently on the CPU and about 5x slower. The tool reads and edits project files, runs tests, builds, Git and shell commands inside the project after asking before touching anything outside it, and exposes an OpenAI-compatible endpoint via `localcode api`. Agentic coding. Local models. On your Mac. uv pip install localcode Upgrade with uv pip install -U localcode . cd your-project localcode To use the running model from another local app, open /server or run localcode api for its OpenAI compatible endpoint. See the Local API guide https://mjwsolo.github.io/localcode/guides/local-api/ . - Reads and edits files in your project - Runs your tests, builds, Git and shell commands inside the project on its own, and asks before touching anything outside it - Searches code by name, content or structure - Scaffolds and launches apps, then checks that they respond - Remembers the task across messages - Mac with Apple Silicon, macOS 13 or newer - 16 GB unified memory or more - Python 3.10 or newer - Disk space for LocalCode and your chosen model download sizes below Choose from 12 models and 169 quant options for coding, image input and typed decisions. Models download from Hugging Face. Downloads run in parallel and can resume if interrupted. Switch with /models . LocalCode recommends the most capable one that fits your Mac's memory and marks it with a star. Speed: Gemma 4, Qwen 3.6, Muse Glimmer and Qwen 3.8 draft tokens with the vendor's own speculative drafter, which downloads with the model. The model server's GPU kernels ship precompiled, so the first start after an install is quicker. | Model | Smallest download | Largest download | |---|---|---| | Gemma 4 12B https://huggingface.co/unsloth/gemma-4-12b-it-GGUF | 4.7 GB | 24.3 GB | | Qwen 3.8 27B https://huggingface.co/unsloth/Qwen3.8-27B-GGUF | 6.2 GB | 54.7 GB | | Inkling 2 Mini 26B-A5B https://huggingface.co/unsloth/Inkling-2-Mini-GGUF | 7.8 GB | 22.9 GB | | North-Mini-Code 30B-A3B https://huggingface.co/unsloth/North-Mini-Code-1.0-GGUF | 9.4 GB | 61.0 GB | | Gemma 4 26B-A4B https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF | 10.4 GB | 51.0 GB | | Qwen 3.6 35B-A3B https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF | 10.5 GB | 69.8 GB | | Flint 40B-A2.5B https://huggingface.co/unsloth/flint-GGUF | 11.4 GB | 79.6 GB | | Muse Glimmer 30B https://huggingface.co/unsloth/Muse-Glimmer-30B-GGUF | 12.4 GB | 57.4 GB | | Gemma 4 Diffusion 26B-A4B https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF | 16.8 GB | 50.5 GB | | OpenJev 27B decisions https://huggingface.co/ggml-org/OpenJev-GGUF | 19.4 GB | 53.8 GB | | Kolibri 1 78B-A3B https://huggingface.co/Hob-forge/Kolibri-1-GGUF | 28.6 GB | 83.1 GB | | DeepSeek V4 Flash 284B-A13B https://huggingface.co/unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF | 82.4 GB | 161.9 GB | What LocalCode adds on top of the stock model server to make local models faster and better at coding. Speed - Drafting: the vendor's own drafter guesses tokens ahead about 2x faster replies on Gemma 4, Qwen 3.6, Qwen 3.8 and Muse Glimmer - Copy drafting: when the model rewrites a file it already read, it copies instead of generating 3.5 to 4x faster on rewrites - Prompt cache reuse: the system prompt stays the same every turn, so the model only reads what is new - Warm-up: the first request's prompt is read while the model loads, so your first message starts sooner - Precompiled GPU kernels: no half-minute shader compile on first start - Threads matched to your Mac's performance cores - Flash attention on by default - Inkling 2: its experts run on the GPU were silently on the CPU, about 5x slower , banded attention and a fused activation op halve its GPU memory and read prompts 20 to 30% faster - Downloads run in parallel, resume if interrupted, and keep going while you use another model Memory - 8-bit KV cache by default: half the memory of full precision, so longer context fits - Inkling 2 uses a 16-bit cache instead: its attention reads it directly, so long prompts read 1.4 to 1.8x faster and it needs 3.3 GB less memory at 128K context - Context sized from what the server reports, so long chats compact before they run out of room - The picker recommends the most capable model that fits your Mac Quality - Hidden thinking off by default, so answers come straight away; a model that gets stuck thinking is recovered instead of failing - Plan panel: the model keeps a task list and is reminded to update it - Stuck detection: a model that makes no progress, or keeps failing the same check, is nudged to change approach and then stopped so it reports what it has - Every release is tested live on every model load, reply, tool calls, images before it ships Inference is local. Three features use the network: model downloads, the web search and web fetch tools, and any MCP servers you add. See Network Boundary https://mjwsolo.github.io/localcode/concepts/network-boundary/ for the full list. Powerful, personal AI should work for everyone, on any device, anywhere. That means running it locally. localcode is a first step. To sponsor localcode, reach out https://github.com/mjwsolo/localcode . See CONTRIBUTING.md https://github.com/mjwsolo/localcode/blob/main/CONTRIBUTING.md . Apache 2.0. See LICENSE https://github.com/mjwsolo/localcode/blob/main/LICENSE .