Sandbox: VS Code devcontainer

Codehamr runs shell commands written by an LLM. A devcontainer keeps every command inside a Linux container you can throw away. Your project files mount in, the model stays on the host's GPU.

Works the same on macOS, Linux, and Windows (where you can also use WSL2 as the sandbox instead).

What you need

VS Code with the Dev Containers extension.
Docker Desktop (macOS, Windows) or Docker Engine (Linux).
Ollama on the host machine with a model pulled. See the Ollama guide or the Windows quickstart.

The layout

Two files inside your project folder:

codehamr-sandbox/
├── .devcontainer/
│   └── devcontainer.json
└── .codehamr/
    └── config.yaml

No Dockerfile, no launch.json, no boilerplate.

1. Devcontainer config

.devcontainer/devcontainer.json

{
    "name": "codehamr-sandbox",
    "image": "mcr.microsoft.com/devcontainers/base:debian",
    "postCreateCommand": "curl -fsSL https://codehamr.com/install.sh | bash",
    "runArgs": ["--add-host=host.docker.internal:host-gateway"]
}

runArgs is needed on Linux. macOS and Windows get host.docker.internal for free, but the line is harmless there.

2. Codehamr config

The one thing different from bare metal: host.docker.internal, not localhost.

.codehamr/config.yaml

active: local

models:
    local:
        llm: qwen3.6:27b
        url: http://host.docker.internal:11434
        key: ""
        context_size: 65536

From inside the container, localhost is the container. host.docker.internal is your host machine.

3. Open it

Ctrl/Cmd + Shift + P → Dev Containers: Reopen in Container. First boot pulls the base image and installs codehamr (~1 min).

Then in the integrated terminal:

codehamr

If something doesn't work

curl -s http://host.docker.internal:11434/v1/models

JSON → recheck url: in your config.
Connection refused → Ollama isn't running on the host.
codehamr: command not found → open a fresh terminal so PATH reloads.

One container, one config, one local model. Done.