orangu 1.0.0

mnemosyne systems is proud to announce orangu 1.0.0 - the Open Source AI-powered coding environment.

orangu is a complete, self-contained AI coding stack - not just a client. Every layer of the pipeline now ships from this one project, written end to end in Rust , and every layer is Open Source under the GNU General Public License v3.0 .

One stack, three programs

Most local AI setups are a patchwork: one tool for the editor, a separate engine for inference, and glue to manage which model is loaded. orangu is the whole stack in one project:

• orangu - the workspace-aware coding environment you drive: the terminal interface, local and Git/forge tools, /review and /auto_review, the knowledge graph, semantic /search, and the context-compression engine
• orangu-coordinator - an optional companion HTTP proxy that starts and stops the server on demand and swaps to whichever model each request needs, so a single-GPU machine can use a different model per role without ever running more than one server
• orangu-server - the inference engine itself: GGUF loading, tokenization, the transformer forward pass, sampling, and request scheduling implemented directly in Rust, with no llama.cpp, no ggml, and no Python

Every layer talks to the next over the OpenAI-compatible API, so the pieces stay cleanly separated - yet they build, ship, install, and run as one.

The new inference server

The headline feature of 1.0.0 is orangu-server, a native, pure-Rust GGUF inference server:

• Model coverage: text-in/text-out GGUF chat, completion, and embedding models across six architecture families - Llama-style, Gemma (dense and routed-expert MoE), Qwen3.5/3.6-MoE, Qwen3.5 dense, Phi-3/Phi-4-mini, and Mistral 3 - read as F32/F16/BF16, Q8_0/Q4_0/Q5_0, the K-quants and the I-quants
• Backends: CPU plus four GPU backends - Vulkan, CUDA, OpenCL, and ROCm - cross-checked in automated tests against the CPU backend’s output
• Serving: OpenAI-compatible /v1/chat/completions, /v1/completions, /v1/models, and /v1/embeddings, alongside native health, props, slots, and metrics endpoints, with continuous batching, prefix caching, flash attention, and durable slot persistence so a resumed conversation does not need to re-prefill
• Workspace file API: eight endpoints operating inside the server’s workspace root, refusing any path that escapes it - the same shared implementation the client’s own file tools use
• Web console: a browser chat UI served from the same binary, with markdown and syntax highlighting, LaTeX, live tokens/second, persistent history, and file attachments - no build step, no external assets
• Model inventory: the same binary manages the machine’s models with list, show, download, delete, prune, suggest, and system

The coordinator has moved into orangu-server, and all GGUF commands now live there as well.

Code review, built in

Code review is orangu’s standout feature, and it runs entirely on your own hardware. orangu turns the terminal into a review workstation for the changes on your current branch - committed work plus what is still in the working tree - measured against the merge base with the default branch. Both reviewers require the branch to be rebased up to date, so you never review against stale code.

/review - interactive review. A full-screen, two-pane view with the file checklist on the right, the selected file’s diff on the left, and your prompt at the bottom:

• Mark each file approved or rejected
• Comment on any diff line under a chosen category - Overall, Code, Security, Memory, Performance, Test Suite, Documentation - plus whole-patch notes
• Ask the connected model about the selected file on demand - focus on error handling, is this thread-safe? - and the exchange joins your chat session for follow-up
• Open any workspace file in your $EDITOR without leaving the view

On exit it writes a category-grouped report with an approve/reject Conclusion, and copies the Markdown to the clipboard.

/auto_review - LLM-driven review. The model reviews the whole change and each file on its own, sorts findings into the same seven categories, marks every file approved or rejected, and renders a final verdict - orangu approves/rejects this patch. It is deliberate about effort:

• File type decides what is scanned: lock files and binary assets are auto-approved with no requests, documentation is reviewed only for the Documentation category, and source files get the full set of checks
• Per-file depth cycles Normal → Deep → Ignore. Deep never truncates the diff, pulls in cross-file context from the workspace’s knowledge graph - who else calls the changed code - and re-checks any rejected finding before accepting it
• A Rigorous Review Rubric combined with confidence scoring (0-100) filters out false positives, hallucinations, and pedantic nitpicks, so only high-confidence findings that meaningfully affect correctness, security, or performance are flagged
• Every finding is pinned to its file:line, and requests are length-capped and tool-free so reviews stay fast and bounded even on modest local models
• After the run you can browse the report, override any verdict, and remove findings before exporting

/auto_review <file> reviews a single file, /auto_review all reviews every Git-tracked file in the project, and immediate and deep combine with either form. Share the result without leaving the terminal: /export review writes a PDF to the workspace root, and /comment 42 with auto review posts it straight onto a GitHub or GitLab issue or pull request. No gh or glab needed.

Because the scheduler runs commands unattended, a line such as 0 6 * * * auto review immediate && export auto review in ~/.orangu/schedule produces a review report each morning with nobody at the keyboard.

Open Source

orangu is licensed under GPL-3.0, and that applies to the entire stack - the environment, the coordinator, and the inference engine. There is no open core, no proprietary engine behind the Open Source front end, and no feature held back for a paid tier. You can read every line that touches your code, from the prompt you type to the matrix multiply that answers it.

The project is developed in the open on GitHub, and 1.0.0 carries contributions from the community across the review workflow, the terminal interface, Git operation error handling, and GPU detection on Apple Silicon. Our thanks to everyone who filed an issue, opened a discussion, or sent a patch.

One vendor

A local AI coding setup normally means assembling parts from several suppliers and owning the integration yourself. orangu 1.0.0 removes that: one project, one language, one license, one installer, one manual, one issue tracker, one release.

The one-liner installer puts the whole stack - orangu, orangu-coordinator, and orangu-server - on the machine in a single step, on Linux, macOS, and Windows:

curl -fsSL https://mnemosyne-systems.github.io/orangu/install.sh | sh

Single vendor, but not single-vendor lock-in. Because every layer speaks the OpenAI-compatible API, you can point orangu at another server, or point another client at orangu-server, whenever you want to. And because nothing leaves your machine, there is no third-party cloud in the loop at all: zero telemetry, no per-token billing, no account, and no Internet connection required once the server and your models are downloaded. That makes orangu a fit for privacy-sensitive, regulated, and air-gapped environments.

Also in this release

• A refreshed terminal interface, now built on ratatui, with updated themes
• Workspace file operations in the server
• Web console improvements: LaTeX support, timestamps, attachments, saved answers, tokens/second, persistent history, and a downloadable debug report
• Batch forward, prefix caching, persistent slots, Q4 and Q8 work, and MoE support in the engine
• GPU memory type detection fixes on Apple Silicon, and a clean refusal of the OpenCL backend on Apple targets rather than a crash
• Error propagation fixes across Git operations, session saving, and the embedding cache
• The complete manual ships with every release as PDF and HTML, and is embedded in the binary for offline reading with /manual

Availability

orangu 1.0.0 is available now for Linux, macOS, and Windows on x86_64 and aarch64, with source and checksums published alongside the binaries.

Commercial support for orangu is available from mnemosyne systems - contact sales .

Be part of the community, and be proud of your contributions ! Try it, star it, fork it, break it, tell us what’s missing.

Thanks for your time !

Overview: https://www.mnemosyne-systems.ai/products/orangu
Repository: https://github.com/mnemosyne-systems/orangu
Release: https://github.com/mnemosyne-systems/orangu/releases/tag/1.0.0