OpenPullRequest

CLI tool that generates PR titles, descriptions, and post-review change summaries using a local or cloud LLM. Supports Ollama, OpenAI, Anthropic, Google Gemini, OpenRouter, DeepSeek, llama.cpp, and MLX.

Built primarily for local-first use — pair it with Ollama and a Gemma model (Google's Gemma family is currently the best fit for this kind of task at a size you can run locally) for a free, fully offline workflow. On Apple Silicon Macs, prefer the -mlx tagged variant of your chosen model (e.g. gemma4:26b-mlx) — it runs on Apple's MLX framework instead of plain GGUF and is noticeably faster on that hardware. Don't want to run anything locally? Google's Gemini API has a generous free tier and is a great low-friction cloud option.

Installation

Recommended

uv tool install openpurr

Or

pipx install openpurr

On first run, the setup wizard creates ~/.openpurr automatically. You can also run it any time:

opo setup

Requirements

• Python 3.11+
• git on PATH
• One of: Ollama running locally, or an API key for a cloud provider

Usage

# Generate a PR title + description from the diff against main
opo

# Diff against a different base branch
opo --base develop

# Summarise changes made since the last N commits (for reviewers)
opo review
opo review --commits 3

Configuration

Settings are stored in ~/.openpurr as flat OPO_KEY=value lines — no sections, no quoting:

OPO_PROVIDER=ollama
OPO_MODEL=gemma4:26b
OPO_API_KEY=
OPO_HOST=http://localhost:11434
OPO_TEMPERATURE=0.0
OPO_KEEP_ALIVE=5m
OPO_BASE=main

Use the config subcommand to inspect and update values without editing the file directly.

# Show all keys with descriptions and current values
opo config describe

# Read a single value
opo config get model

# Update a value
opo config set provider ollama
opo config set model gemma4:26b
opo config set api_key sk-...
opo config set keep_alive 0s

Custom system prompts

The prompts opo sends to the LLM aren't fixed — everything after a lone --- line in ~/.openpurr is free text, not OPO_KEY=value config, so it can hold Markdown, headers, = signs, anything. Add an INIT PROMPT: section to override the prompt behind opo (title + description) and/or a REVIEW PROMPT: section to override the one behind opo review. Either section is optional; whichever is missing keeps the built-in default.

OPO_PROVIDER=ollama
OPO_MODEL=gemma4:26b
OPO_API_KEY=
OPO_HOST=http://localhost:11434
OPO_TEMPERATURE=0.0
OPO_KEEP_ALIVE=5m
OPO_BASE=main

---
INIT PROMPT:
You are a Senior Principal Engineer. Write a terse PR title and description in
Conventional Commits style. No emoji, no checklists.

REVIEW PROMPT:
Summarize what changed since the last review in one short paragraph.

opo config set/opo config describe only ever touch the OPO_KEY=value part above the --- line — the prompt section is left exactly as you wrote it, and opo config describe tells you which of init/review are currently overridden.

Configuration keys

There's no hardcoded default model: opo setup always has you pick one (from a live-fetched list, or typed manually), and opo/opo review will error out with a pointer back to opo setup if model is ever left blank (e.g. after hand-editing the file).

Example: switch to Google Gemini (free tier)

opo config set provider gemini
opo config set model <your-chosen-model>
opo config set api_key <your-gemini-api-key>

Model names aren't hardcoded anywhere in this README on purpose — providers retire and rename models constantly. Run opo setup or opo models --provider gemini to pick from what's actually available right now.

Example: unload Ollama from VRAM after each request

opo config set keep_alive 0s

Models

List available models for the current or a specified provider:

opo models
opo models --provider anthropic

Model lists are always fetched live from the provider — Ollama's /api/tags, OpenAI/Anthropic/Gemini/OpenRouter/DeepSeek/llama.cpp/MLX's models.list() (or public models endpoint) — never a hardcoded/curated list baked into the tool. If the fetch fails (offline, bad key, local server not running) the command prints an empty result instead of erroring; check connectivity/credentials, or pull a model directly with ollama pull <model>.