OpenCrust
Self-hosted AI assistant platform for Linux, macOS and Windows, with Ollama local models, cloud APIs, messaging integrations and an MIT license.
OpenCrust is a self-hosted AI assistant platform for people who want personal agents available through their chat apps. It runs on Linux, macOS and Windows as a compact Rust application and is open source under the MIT license. A browser interface and terminal chat provide direct access alongside Telegram, Discord, Slack, WhatsApp, LINE, WeChat and macOS iMessage.
You choose where inference runs. Ollama handles local models, while vLLM connects models on your own server. Claude, OpenAI and other OpenAI-compatible services are also supported; requests to those providers leave your machine. OpenCrust stores conversation memory and ingested documents locally in SQLite. Semantic document search uses an embedding provider, with Cohere as the default; keyword search works without one.
Named agents can delegate tasks to one another, each with its own model, personality, skills and permitted tools. Agents can work with files, execute shell commands, search the web and carry out scheduled tasks. MCP connects them to additional tools and data sources. Document search retrieves relevant passages from PDFs, text files and source code, and answers cite the documents they use.
The assistant can turn repeated workflows into reusable skills and revise those skills after use. It accepts agentskills.io-compatible skills and scans them for prompt-injection patterns before installation. Voice input and replies can use local Whisper and self-hosted Kokoro, or OpenAI's cloud services. Access controls include encrypted credential storage, WebSocket pairing, per-channel authorization and tool restrictions; optional WebAssembly plugins run in a sandbox.
