AgentArk

AgentArk is a self-hosted AI agent runtime for people who want a personal assistant that remembers preferences and carries out ongoing work under their control. It runs through Docker on macOS, Windows and Linux, with a browser interface for chat, tasks and action review. It's free and open source under Apache 2.0.

The assistant can schedule routines, monitor inboxes, draft replies and follow up across messaging channels. It also builds apps and deploys them with public URLs. Gmail, Google Calendar, Telegram, WhatsApp and Slack integrations connect those tasks to existing accounts; MCP servers and webhooks extend what agents can use. Web research can use a self-hosted SearXNG backend.

Memory persists across conversations. Accepted work, corrections and tool results feed into proposed changes to memory, prompts and task routing. Users can review those improvements, inspect stored facts and roll memory back. The runtime also reduces bulky browser output and logs before they consume model context.

Docker and Wasmtime isolate execution, while permission checks send actions that aren't pre-authorized to an approval queue. Audit trails record agent activity. Conversation history, memory and integration tokens stay in local storage, and AgentArk encrypts secrets. It doesn't require an AgentArk account or managed backend.

Ollama supports local model use. Anthropic, OpenAI, Gemini and Groq are optional cloud providers that receive model requests directly when selected. Web search and connected services use external networks. AgentArk currently keeps all work in one shared workspace.