OpenBMB/PilotDeck

Task-oriented AI Agent productivity platform

What it solves

PilotDeck is a task-oriented AI agent productivity platform designed for long-running, multi-project work. It addresses common agent limitations such as "black-box" memory (where users cannot see or edit what the AI remembers), high token costs when using flagship models for simple tasks, and the inability of agents to work autonomously in the background while the user is away.

How it works

The platform is organized around "WorkSpaces," which isolate files, memory, and skills for each project to prevent context pollution. It employs three core mechanisms:

  • White-box Memory: Makes memory generation, extraction, and retrieval visible and editable, allowing users to fix incorrect memories and rollback states via a "Dream Mode."
  • Smart Routing: Automatically detects task difficulty and routes complex requests to flagship models (e.g., Claude 3.5 Sonnet) and simpler tasks to lighter models to reduce costs.
  • Always-on Execution: Enables agents to proactively discover tasks, monitor progress, and deliver results as local files without constant user prompting.

It natively supports the Model Context Protocol (MCP) and provides a Web UI, CLI, and IM integration.

Who it’s for

It is built for users engaged in complex, multi-task productivity work, such as AI engineering, mini-game development, and social media operations, who need a traceable and cost-effective agent system.

Highlights

  • WorkSpace Isolation: Dedicated file systems and memory stores per project.
  • Traceable Memory: End-to-end visibility and manual editing of memory entries.
  • Cost Optimization: Smart routing that can reduce token spend by up to 70% on specific workloads.
  • Background Autonomy: Ability to run long-horizon monitors and land deliverables as files on disk.
  • Extensible Architecture: Native MCP support and a plugin system for custom tools, skills, and memory stores.

Related

  • Project
    hotovo/aider-desk

    An open-source agentic platform and orchestration layer for AI-assisted development that provides professional engineers with transparent control over AI coding assistants.

  • Project
    codejunkie99/agentic-stack

    agentic‑stack is an open‑source toolbox that gives a portable “brain” (memory, skills, protocols) to many code‑generation agents (Claude Code, Cursor, Copilot CLI, Gemini CLI, etc.). The brain lives in a hidden `.agent/` folder, enabling agents to retain preferences, lessons, and reusable skills across tool switches. It provides bounded agentic loops, a review‑by‑human protocol, a local dashboard for token‑cost/KPI monitoring, and export utilities that turn approved runs into trace records or training‑ready JSONL. Installable via Homebrew or a simple script, it supports optional external long‑term memory (the `brain` server) and a beta full‑text memory search.

  • Project
    ayuayue/PiDeck

    An open-source desktop workbench for managing multiple pi and DeepSeek Harness (DSH) AI coding agent sessions across local projects.

  • Project
    TencentCloud/Octop

    Octop is an open‑source, self‑hosted AI assistant platform. It runs a single FastAPI‑based process that serves a React dashboard, a CLI, and bots for Feishu, DingTalk, QQ, Discord, WeCom, etc. Users can create multiple agents with distinct MBTI‑style personalities, each with its own persistent memory and configurable LLM provider (OpenAI‑compatible, DashScope, Ollama, …). All data stays locally under `~/.octop/`, and the system includes guardrails, JWT isolation, browser/terminal automation, and a bidirectional ACP protocol for IDE integration. Installation is via a one‑line script, PyPI, or Docker; the project is MIT‑licensed and actively maintained.

  • Project
    CyberStrikeus/CyberStrike

    CyberStrike is an open‑source AI agent that turns any LLM (Claude, GPT, Gemini, etc.) into an autonomous penetration‑testing platform. It adds a security‑focused “intelligence layer”, ships 13+ specialised agents and 7,600+ signed security‑skill files, supports 150+ AI providers (including offline models), and can orchestrate real tools locally or on remote Bolt servers. A TUI, optional web UI, and built‑in HackBrowser complete the workflow.