A self-hosted AI platform where a master LLM orchestrates workers, manages SSH machines, connects MCP tools, and runs autonomously 24/7 with the heartbeat agent. Your models, your rules, your server.
A master LLM plans the work and delegates subtasks to specialized workers in parallel. Cheaper, faster, better.
Register your servers. The agent runs commands, streams output live, uploads and downloads files — all from chat.
Connect any MCP-compatible tool server. Workers get scoped access to exactly the tools they need.
Each user gets an isolated file system. Upload, browse, drag into chat. The agent reads and writes files for you.
Long conversations are automatically summarized when the context window fills up. Nothing is lost, history is always retrievable.
Your server, your data. Secrets encrypted at rest, per-user isolation, configurable data retention. You control what stays and what goes.
Route sensitive data to local models only. Per-model safety tags, audit-ready memory system, and data export tools — designed for teams that care about how their data is handled.
The heartbeat agent keeps your research moving while you sleep. It reads state, decides what's next, and drives the conversation forward — 24/7.
Delegation plans auto-execute after a configurable countdown. Freeze to edit, or let the timer run. Autonomous loops without babysitting.
Workspace editing, SSH machines, web search, code analysis, persistent memory, plan tracking, file management — all built in.
Define behavioral rules as plain Markdown files. Link them per conversation. The heartbeat agent gets its own dedicated instruction file.
Every message is a typed component — text, code, plans, traces, SSH output, heartbeat messages, todos, and more. Extensible with one file.
Ask anything — build a feature, analyze data, deploy to a server, research a topic.
Your master LLM breaks the request into subtasks, picks the best worker for each one, and shows you the plan.
Reassign workers, edit instructions, add or remove tasks. You're always in control.
Each worker gets only the context it needs — trimmed, scoped, budget-capped. No wasted tokens.
Results are collected and the master produces a unified response. A trace shows exactly what happened — LLMs used, tools called, tokens spent.
Modern stack, no magic, fully open.