Repolode
Submit
dwim-sh

dwim-sh/dwim

A coding agent that runs a local 27B model on your own GPU, with no API key and no data leaving your machine.

View on GitHubWebsite
Stars
27
Forks
0
Open issues
12
Last commit
4 hours ago

What is it?

dwim is a coding agent written in Rust that bundles the agent, the model and the inference engine into one program. The default model is Bonsai 2 27B, a Qwen-based model compressed to about 6 GB with ternary weights, and it is downloaded on first use. Once installed, no API key or external server is needed, and the code stays on the local machine.

Run dwim in a project directory to open an interactive shell. You type a request, and the model reads files, runs shell commands, reads their output and continues until it answers. A single prompt can also be passed on the command line for one-off questions. Inference uses Metal on macOS and Vulkan on Linux.

dwim is experimental. It runs shell commands without asking for confirmation or applying a time limit, so it is designed to run inside a container or virtual machine, and the project recommends using it only on code that is committed or backed up.

Key features

  • 01Runs a local coding agent that executes commands and reads their output.
  • 02Ships the Bonsai 2 27B model, compressed to about 6 GB with ternary weights.
  • 03Uses Metal on macOS and Vulkan on Linux for local inference.
  • 04Provides an interactive shell and one-shot prompts from the command line.
  • 05Requires no API key or external server.

Use cases

  • Exploring and explaining an unfamiliar codebase without sending code to a third party.
  • Running a coding assistant on an offline or air-gapped machine with a capable GPU.
  • Experimenting with a local coding agent inside a disposable container or VM.

Quick start

  1. 1

    Run the official install script from the README

  2. 2

    Or install the development version with Cargo

    cargo install --git https://github.com/dwim-sh/dwim
  3. 3

    Run dwim in a project directory; the model (about 6 GB) downloads on first use

    dwim

Things to know

  • The project is experimental and may have bugs and breaking changes between commits.
  • The model runs shell commands without asking first, so it should be used in a sandbox on committed or backed-up code.
  • It needs a local GPU and a roughly 6 GB model download on first use.

In collections

AI coding agentsCoding with local LLMsAI coding assistants & agentsRun LLMs locallyCLI & terminal toolsPrivacy & local-first tools

Topics

#coding-agent#local-llm#rust#inference#privacy#cli

FAQ

What is dwim?
A coding agent that runs a local 27B model on your own GPU, with no API key and no data leaving your machine. dwim is a coding agent written in Rust that bundles the agent, the model and the inference engine into one program. The default model is Bonsai 2 27B, a Qwen-based model compressed to about 6 GB with ternary weights, and it is downloaded on first use. Once installed, no API key or external server is needed, and the code stays on the local machine.
Is dwim free to use? What license does it have?
dwim is open source under the MIT license. Check the license terms before using it in commercial products.
Is dwim still maintained?
Actively maintained. The last commit to dwim was on Nov 7, 2026.
What can I use dwim for?
Exploring and explaining an unfamiliar codebase without sending code to a third party. Running a coding assistant on an offline or air-gapped machine with a capable GPU. Experimenting with a local coding agent inside a disposable container or VM.
How do I install dwim?
Run the official install script from the README Or install the development version with Cargo: cargo install --git https://github.com/dwim-sh/dwim Run dwim in a project directory; the model (about 6 GB) downloads on first use: dwim
Where can I get dwim?
The source code and installation instructions are on GitHub: https://github.com/dwim-sh/dwim

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects