orangu

Why orangu

The private,
Git-centric companion.

orangu is the lean, private, Git-centric coding companion for the terminal — built for developers who run their own models and want a tightly integrated review workflow without sending a single line of code to the cloud.

100% local and private

Zero telemetry. After the model is downloaded, no internet is needed — it runs happily in air-gapped environments.

Built-in code review

An interactive two-pane reviewer (/review) and a category-by-category LLM auto-reviewer (/auto_review). orangu's standout feature.

The whole Git loop in the prompt

Branch, commit, rebase, squash, cherry-pick, stash, push, and GitHub/GitLab PRs, comments, and issues — all from the terminal.

Pure Rust, top to bottom

Quick startup, no runtime to install, its all Rust based.

## issue-85-prune...origin/issue-85-prune

> /auto_review

Working @ 38.2 t/s (4s)

The prune_sessions_output function looks solid.

One suggestion: the OlderThan match arm uses

saturating_sub correctly — consider

returning the removed session count

alongside the UUID list for better UX.

> /commit "[#85] tool: /prune"

committed on issue-85-prune

> /push

pushed to origin/issue-85-prune

The stack

Three cooperating programs,
one Pure-Rust project.

orangu is a complete, self-contained AI coding stack — not just a client. Every layer is open source under GPL v3, and there is no proprietary backend to call home to. Two further binaries ship in the same archive: the benchmarking tool orangu-bench, and orangu-gguf, which builds the model files the rest of the stack runs on.

Diagram of the orangu stack: you talk to orangu, the coding environment; orangu talks over OpenAI-compatible HTTP to the optional orangu-coordinator, which starts and stops orangu-server and swaps the model per role; orangu-server runs the GGUF engine on CPU, Vulkan, Metal, CUDA, ROCm, or OpenCL. — One native stack — fully local, fully private, entirely your own.

orangu

The coding environment: terminal UI, workspace-aware local tools, the full Git and GitHub/GitLab loop, code review, knowledge graph, and semantic search.
orangu-coordinator Optional

An on-demand model manager. It starts and stops the server for you and swaps to whichever model each role needs — chat, review, and the rest — so a single-GPU machine never has to run several servers at once.
orangu-server

A native pure-Rust GGUF inference engine: model loading, tokenization, the transformer forward pass, sampling, and request scheduling — with no llama.cpp, no ggml, and no Python. It serves OpenAI-compatible endpoints and a built-in web console.

CPUVulkanMetalCUDAROCmOpenCLNPU
orangu-bench Tool

A benchmarking tool rather than part of the stack. It measures decode and prefill throughput of any OpenAI-compatible server over HTTP and charts it over time, so a backend change or a new quantization can be judged on your own hardware instead of someone else's numbers. Runs from the command line or the web console.
orangu-gguf Tool

The model builder, new in 1.3.0. It pretrains a model from scratch out of a manifest of permissively-licensed repositories — cloning them, training a tokenizer, packing the text into tokens, and writing the result as a GGUF file orangu-server loads directly — and it rewrites an existing model at a smaller weight format, from q6_k down to q2_k.

Code review & auto review

Your terminal is now a
code-review workstation.

This is orangu's standout feature. Review the changes on your branch — committed and uncommitted, measured against the merge base with your default branch — two ways: an interactive reviewer you drive, and a fully automated one the model runs end to end. Both require an up-to-date rebase, so you never review against stale code.

/review

Interactive review

Full-screen two-pane view — file checklist beside the selected file's diff
Mark each file approved or rejected; comment on any line under seven categories
Ask the model about a file on demand — the exchange joins your chat session
Open any file in $EDITOR without leaving the view
Exit writes a category-grouped report with an approve/reject conclusion, copied to your clipboard. No gh/glab needed

/auto_review

LLM-driven review

The model reviews the whole change and each file on its own, sorting findings into seven categories
Renders a final orangu approves / rejects this patch verdict
Smart about effort — lock files and binaries auto-approve, docs get doc checks, source gets the full set
Every finding pinned to its file:line; requests are length-capped and tool-free to stay fast on local models
Browse the report, override any verdict, then /export review or post with /comment

Features

Everything, in the terminal.

Private

Fully local

Runs on orangu-server or any other OpenAI-compatible server. Once models are downloaded, nothing leaves your machine.

Review

Code review built in

Interactive /review and automated /auto_review — the standout in-terminal review workflow.

Tools

Workspace-aware

Read, edit, list, /grep, fetch, and run shell commands — all scoped to your workspace.

Workspaces

Projects as tabs

Open several projects at once, each a tab with its own session, scrollback, queue, and history.

Skills

Agent Skills

Drop a SKILL.md into your repo; orangu discloses it progressively and loads it on demand.

Terminal

Built for the prompt

Ghost-text hints, Tab completion, live tokens/sec, queued commands, and Markdown with syntax highlighting.

Automation

Workflow files

A YAML file of jobs, variables, functions, and approvals — the same slash commands, run unattended across several workspaces.

Models

Build your own model

orangu-gguf pretrains a model from permissively-licensed sources, or rewrites one you have at a smaller weight format.

Patches

Review into fixes

/create_patch turns a review report into the change it asked for — and unpicks a conflicted merge, rebase, or cherry-pick.

What's new

The latest release.

orangu 1.3.0 Latest September 9, 2026

Train your own model — orangu-gguf pretrains a model from a manifest of permissively-licensed repositories and writes it straight out as GGUF, or rewrites one you already have at a smaller weight format.
Workflows — A YAML file describes jobs, variables, functions, approvals, loops, and control flow, and runs the same slash commands you type. --dry-run checks the whole file before anything executes.
NPU support — A seventh backend joins CPU, Vulkan, Metal, CUDA, ROCm, and OpenCL: Rockchip’s RKNPU on boards like the RK3588.
/create_patch — Turns a review report into the fixes it asked for, and resolves the conflicts a stopped merge, rebase, or cherry-pick leaves behind.
New models — Qwen 3.8-Flash-Next with multi-token prediction, GLM 5.3, and Ornith 1.5.
Model list, sorted — The server’s model list sorts, with a refresh-all option, and LAST_USED is tracked in plain orangu-server too.

Also in 1.3.0: Windows fixes, dependency updates, and further performance work across the engine. The complete manual ships as PDF and HTML with the cheat sheet, and is embedded in the binary for /manual.

See all releases →

> /status

## feature/new-api...origin/feature/new-api

M src/config.rs

M src/bin/orangu/init.rs

> /diff

+ pub line_endings: String,

> /squash

squashed 3 commits into 1

> /pull_request

PR #96 created

Git integration

Your entire Git workflow,
from one prompt.

Commit, push, rebase, squash, review, and open pull requests, all without leaving the terminal. orangu understands your branch, your history, and your forge.

GitHub and GitLab

Create PRs, close issues, add review comments via gh or glab.

Full-screen review mode

Two-pane diff view with AI feedback, line annotations, and keyboard navigation.

Auto-sync on startup

Fast-forwards your default branch against origin silently every session.

orangu vs. a cloud coding assistant

A deliberate trade.

orangu makes a deliberate trade: a focused, offline-first, Git-centric terminal experience instead of a broad cloud platform. If you run your own models and care about privacy, that trade is the whole point.

	orangu	Typical cloud coding assistant
Where your code goes	Stays on your machine — zero telemetry	Sent to a third-party provider
Offline use	First-class; only the initial model download needs a network	Generally requires connectivity
Models	GGUF models you own, served by the bundled `orangu-server` — or any other OpenAI-compatible server	Vendor-hosted models, usually behind API keys
Cost	Free to run against models you host	Per-token / subscription billing
Footprint	Native Rust binaries, fast start, no runtime, no Python	Editor/cloud service + account
Code review	Built-in interactive and LLM auto review in the terminal	Usually delegated to the hosting platform
Git workflow	Full Git + GitHub/GitLab loop from the prompt	Varies; often browser-based
Privacy posture	Suited to regulated / air-gapped environments	Depends on the provider's data policy