OpenBMB/PilotDeck
An operating system for agents that won't shut up (in a good way)
PilotDeck keeps AI agents running, routed, and memory-honest across long projects instead of one-shot chats.
Velocity · 7d
+179
★ / day
Trend
→steady
star history
What it does
PilotDeck is an open-source “agent operating system” built around WorkSpaces — isolated project containers that each get their own files, memory, and skills. It runs as a persistent server with Web, CLI, and chat-app front ends, and it natively speaks MCP (Model Context Protocol). The pitch is moving from “ask once, answer once” to agents that keep working while you’re away.
The interesting bit
The project treats memory as white-box infrastructure: you can inspect, edit, pin, or roll back individual memory entries per WorkSpace, and a “Dream Mode” compresses memory during idle time with one-click undo. That’s paired with Smart Routing, which auto-detects task difficulty and demotes grunt work to cheaper models. The README claims ~70% cost savings on social-media workflows and a 1/6 cost reduction on hard tasks while matching or beating single frontier-model performance.
Key highlights
- WorkSpace isolation: Files, memory, and skills are scoped per project; no global context pollution.
- White-box memory: Full visibility into what the agent stored, when, and why; editable and reversible.
- Smart Routing: Auto-switches between flagship and lightweight models by task difficulty; benchmarks provided for social-media and complex multi-step workloads.
- Always-on execution: Agents discover tasks, run long-horizon monitors, and write deliverables to disk with summary reports.
- MCP-native: Built on the Model Context Protocol; consistent behavior across Web, CLI, and IM front ends.
Caveats
- The README is heavy on vision and light on architecture specifics; it’s unclear how much is novel runtime versus orchestration glue.
- Benchmarks are self-reported by the research team (Tsinghua/OpenBMB) with no independent validation cited.
- One-line installer is macOS/Linux only; Windows users are left to source builds.
Verdict
Worth a look if you’re burning API budget on agents and need per-project isolation with audit trails. Skip it if you want a lightweight library — this is a full platform with opinions about how agents should live and remember.