ORCA

The open-source autonomous AI agent for your desktop. Bring any model — or none.

Installation · Documentation · Providers · API · Changelog · فارسی

English · Русский · 中文 · فارسی · Documentation site

ORCA is a desktop agent that plans, writes code, runs it, reads the errors and fixes them — on your machine, in your files. It ships with built-in models that work out of the box, and connects to 200+ providers or any OpenAI‑compatible endpoint with your own key.

• Autonomous by default. 46 real tools: shell, Node/Python, file editing, web search, HTTP, headless-browser checks, Office documents, media, OCR, image and video generation, background processes, sub‑agents. Build requests are built first, with sensible defaults — a question is asked only when something genuinely blocks the work (a bot token), and only after the code exists.
• Any model. Built-in models with automatic failover, plus your own providers (cloud or local). Keys never leave your device.
• Builds whole sites. The model writes the content — texts, items, prices, FAQ, testimonials — and scaffold_site renders it into a finished, modern multi-page site in one step (design tokens, dark/light, hero, cards, stats, catalog with filters and cart, validated forms, RTL when needed), then every page is verified in a headless browser. No placeholders, no half-built pages.
• Runs what it builds. Bots, dev servers and workers start as background processes that survive across turns, with a live log, detected ports and a Stop button in the Processes panel. Image and video generation plug into any Images API / Videos API provider (OpenAI, Routeway, Together, xAI, Sora, Replicate, fal.ai) with a Test button in Settings.
• Knows what it can do. Ask "what can you do?" and the answer is generated from a live capability check of your installation — connected models, present binaries, enabled engines — not from a canned list.
• Decision engine (optional). Plug in a System One model and ORCA gets calibrated judgments in ~0.2 s: smarter auto-routing, semantic command-risk review, verification of every answer (garbled, promised-but-not-done, wrong language), re-ranked web results.
• Three ways to chat. Direct, side-by-side comparison, or blind battle with a local leaderboard.
• Checkpoints and diffs. Every file change is snapshotted; restore any step with one click.
• Built for long work. Output caps are continued transparently, long pastes become files, stopped answers can be continued, and pins, notes and branches keep long chats on track.
• Four languages. English, Русский, 中文, فارسی — with matching typography.
• Private. No accounts, no telemetry. Everything is stored in a local data folder you control.

#Installation

Download the build for your system from Releases:

Builds are not code-signed; SHA256SUMS is attached to every release. The app updates itself: it checks the release channel, downloads the package for your OS/CPU in the background (resumable, mirrors when GitHub is blocked), verifies the SHA-256 and installs it on restart — no browser needed (Settings → Updates). Installed Windows builds update through a silent installer, the portable exe replaces its own file, and packaged macOS/Linux builds swap themselves in place.

From source (Windows, macOS, Linux):

git clone https://github.com/Nethyric/orca.git
cd orca
npm install
npm start          # desktop app
npm run web        # or: browser UI at http://localhost:7860

#MCP (Model Context Protocol)

ORCA ships a minimal, dependency-free MCP client: connect external tool servers from Settings → MCP (or paste a Claude Desktop mcpServers config as-is). Both transports are supported — stdio (local subprocess) and Streamable HTTP (with a legacy SSE fallback). Servers spawn lazily and reconnect quietly; every result is treated as untrusted content, tool manifests are fingerprinted to detect silent redefinition, and each tool can be toggled per server. Use compact mode to keep a server's schemas out of the context until the agent discovers them on demand.

#Providers

Open Settings → Models → Add provider, pick a provider from the catalog (or Custom for any OpenAI-compatible URL), paste a key, press Discover to list its models, Test, Save. Keys are stored only in your local config.json.

Anthropic-style and OpenAI-style APIs are both supported, as are local servers (Ollama, LM Studio, vLLM, llama.cpp). Full guide: docs/providers.md.

#Documentation

#Building ORCA

npm run fetch-bins   # ffmpeg, yt-dlp, OCR data (not in git)
npm run dist:win     # release/ORCA-Agent-<version>-win-x64.zip
npm run dist:mac     # release/ORCA-Agent-<version>-mac-{x64,arm64}.{zip,dmg}   (on macOS)
npm run dist:linux   # release/ORCA-Agent-<version>-linux-x64.{AppImage,tar.gz,deb}

Releases are built by GitHub Actions on every v* tag. Maintainers: see docs/setup.md#maintainers for the release and key-rotation workflows.

#Contributing

Issues and pull requests are welcome. Please read SECURITY.md before reporting anything key-related, and open a bug report with the ORCA version and the log excerpt.

#License

MIT © Nethyric