PilotDeck is an agent operating system organized around the WorkSpace. Each project gets its own file system, memory store, and skill set, so parallel projects do not share a context pool, retrieval has a bounded scope, and skills accrete per project. It speaks the Model Context Protocol (MCP) natively and behaves consistently across its Web, CLI, and instant-messaging front-ends.
Memory is white-box: generation, extraction, storage, and retrieval are visible end to end, individual entries can be viewed, edited, deleted, or pinned, and each entry records what was stored, when, and in which WorkSpace. A Dream Mode pass consolidates memory during idle windows and supports one-click rollback to the prior state, so a compression pass does not destroy the original.
Smart Routing detects task difficulty and sends planning checkpoints to a flagship model while demoting simpler steps such as polishing and layout to a lighter sub-agent. Two reported measurements: a social-media operations workload cost $2.83 with routing on against $12.58 with a single flagship model for both main and sub agents, and a seven-task benchmark covering multilingual podcast distribution, multi-source data reports, domain literature review, and codebase architecture docs scored as follows.
| Setting | Score | Cost |
|---|---|---|
| MiniMax-M2.7 single-agent | 37.1 | $1.90 |
| Claude Sonnet 4.6 single-agent | 69.1 | $18.36 |
| Sonnet 4.6 (main) + MiniMax-M2.7 (sub) | 70.6 | $3.15 |
An always-on mode keeps work moving after you sign off: the agent discovers candidate tasks, runs long-horizon monitors, and lands deliverables as local files with a summary report. The plugin architecture keeps a boundary between the open-source core and customization, with a plugin.json registering MCP servers, custom tools and skills, lifecycle hooks such as PreToolUse and UserPromptSubmit, or a replacement memory store provider.
Features
- WorkSpace isolation: per-project file system, memory store, and skill set, so parallel work does not cross-contaminate
- White-box memory: every entry is visible, editable, deletable, and pinnable, with generation through retrieval auditable
- Dream Mode: idle-window memory consolidation with one-click rollback
- Smart Routing: task difficulty detection routes complex calls to a flagship model and simple ones to a lighter sub-agent
- Always-on execution: background task discovery, long-horizon monitoring, and deliverables written to disk with a report
- Provider choice: OpenAI, Anthropic, native Google Gemini, DeepSeek, Qwen, Kimi, MiniMax, and other OpenAI-compatible endpoints, configured in a YAML file or through onboarding in the Web UI
- Local models: Ollama is configured through its OpenAI-compatible endpoint with no API key
- Web UI: WorkSpace management, memory editing, and a view of multi-agent collaboration
- Plugin protocol: MCP servers, tools and skills, lifecycle hooks, and custom memory providers declared in plugin.json
- Cost tracking: token spend is tracked per task, so background runs can be costed
