Patchforge

This repo now ships an installable CLI that wraps:

• llama-server for a local OpenAI-compatible endpoint
• aider-chat for file editing against that local model

The binary name is patchforge.

Install

Install the package into a tool environment:

uv tool install .

Or into a project virtualenv:

python3 -m venv .venv
./.venv/bin/pip install -e .

aider-chat is installed as a package dependency. Then bootstrap the native side:

patchforge install

patchforge install chooses the best local installation path it can find:

• Reuse an existing llama-server if one is already on PATH
• Otherwise, on macOS, prefer Homebrew and install llama.cpp
• After that, prefetch the default GGUF models into the llama.cpp cache

Usage

Start the local model server in the background:

patchforge start

Check whether it is up:

patchforge status

Run Aider against that endpoint:

patchforge aider --yes-always --message "Create hello.txt with a short greeting."

You can also let the CLI ensure the server is running first:

patchforge aider --ensure-server --yes-always --message "Create hello.txt with a short greeting."

Stop the managed background server:

patchforge stop

Inspect the local /v1/models endpoint:

patchforge models

Defaults

• Host: 127.0.0.1
• Port: 8091
• Model alias: gemma-local
• Default models: cached gemma-2-9b-it first, then cached gemma-4-E4B-it
• Runtime state: .patchforge/ under the current project

Overrides

These environment variables are supported:

LLAMA_HOST=127.0.0.1
LLAMA_PORT=8095
LLAMA_MODEL_ALIAS=my-local-model
LLAMA_MODEL_PATH=/absolute/path/to/model.gguf
LLAMA_CTX_SIZE=8192
LLAMA_PARALLEL=1
OPENAI_API_KEY=sk-local

You can also pass the same values as CLI flags such as --port, --model-alias, and --model-path.

If you want to force the Homebrew path explicitly:

patchforge install --installer brew --force-install