[!WARNING] dwim is experimental software. Expect bugs, rough edges, and changes that break things from one commit to the next. The model runs shell commands on your machine without asking first, so use it on code you have committed or backed up.

Installation

To install the latest release:

curl -fsSL https://dwim.sh/install.sh | sh

To install the latest development version:

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

Or, from a checkout, with the manual page:

make install                     # in ~/.local
make install PREFIX=/usr/local   # system-wide, which may need sudo

Either way, the model (16 GB) is downloaded on first use into your Hugging Face cache, where hf download and other tools would put it, and used from there if one of them already did. dwim --download fetches it ahead of time.

Features

• Fully local coding agent. Runs commands, reads their output, and keeps going.
• Qwen3.8-27B. At four bits, 16 GB, on a machine with 32 GiB of memory or more, decoding several tokens at a time with the drafter its file ships.
• Fast local inference. Metal on macOS, Vulkan on Linux.
• Built-in local inference server. The model is loaded once, in a dwim server that every dwim you run connects to. The first one starts it in the background, and it exits when idle. Sessions share the weights, and each has its own context on the GPU, so you can run several at once.
• Bring your own sandbox. dwim expects to run inside a sandbox, such as a container or a virtual machine, and does not confine the model itself. Commands run without confirmation and without a time limit, until they finish.

Getting started

Run dwim in a project directory. It opens a shell: type what you want, and the model reads files, runs commands, and answers.

Run dwim in another terminal and it connects to the same server, ready in a moment instead of loading the model again. Two sessions fit at once by default, each with its own context; --sessions sets how many, and /server in the shell shows, stops, or restarts the server.

To ask one thing and exit, give the prompt on the command line:

dwim "Explain this codebase to me in one sentence."

The manual page has the options, the shell's commands and keys, and the files it reads:

man dwim

Using dwim as a sub-agent

dwim also makes a good sub-agent: your main agent plans the change, and dwim writes the code. The dwim skills teach an agent such as Claude Code or Codex to delegate to it:

npx skills add dwim-sh/skills

Then, in the agent, ask for the change "with dwim":

1. /dwim plan reads the code, decides the design, and writes one self-contained task file per commit under .dwim/plans/<name>/, with a PLAN.md that holds the goal, the design, the checks, and the steps as a checklist. It runs nothing: you read the plan and edit what you want.
2. /dwim do runs the plan step by step: it hands each task file to dwim, watches the run, reviews the change as it would a pull request, fixes and verifies it, commits it with Generated-by: dwim, ticks the step off, and goes on to the next until a step fails review or none remain. /dwim do next or /dwim do 3 runs just one step, and /dwim status shows where the plan stands.

The main agent spends its tokens on the decisions and the review, and dwim, on your own GPU, on the code.

License

MIT.