Neura Market
nikhilkagita04/continuum logo

nikhilkagita04/continuum

Free

把整个桌面上下文（看过、读过、输入过、听到的一切）汇成单一 MCP，供 Agent 理解你的工作方式。本地优先、设备端运行：对焦点窗口做 OCR、可选会议转录，并提炼长期偏好，让 Claude/Cursor 等随时知道“你刚才在某 App 里做了什么”。

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)

About nikhilkagita04/continuum

Continuum is an open-source, local-first MCP server that captures your entire desktop context—everything you’ve seen, read, typed, and heard—and serves it to AI agents. It uses on-device OCR on the focused window, optional audio transcription, and keyboard capture to build a rich, deduplicated timeline. A scheduled 'dreaming' pass consolidates raw activity into structured memory files (projects, preferences, people, taste, decisions) that agents read to understand you. Every recalled moment includes a citation (source app, owner, date) so agents attribute correctly. The system is designed for privacy: memory lives on-device, agents have read-only access, and an append-only audit ledger tracks exactly what data leaves the machine with a build-failing check. Optional encryption at rest is available. Local retrieval, capture, and preferences are free; deep memory (dreaming) requires an external model API key.

Key Features

On-device screen OCR on focused window with content deduplication
Cited dated recall with source app, owner, and calendar date
Out-of-band dreaming pass consolidates raw activity into small memory files (who you are, projects, people, taste, decisions)
Standing preferences (e.g., 'be concise') automatically apply to all agents
Local-first, read-only access for agents (memory is not writable by agents)
Append-only audit ledger of all data leaving the machine with a build-failing check
Optional encryption at rest for memory store
Hybrid lexical + local embeddings + RRF retrieval
Works with any MCP-compatible agent (Claude Desktop, Claude Code, Cursor, Codex)

Pros & Cons

Pros
  • Privacy-preserving: on-device, local-first, audit log of data egress
  • Provides structured, cited memory (projects, preferences) to agents
  • Open-source and free for local capture, retrieval, and preferences
  • Reduces friction of re-pasting and re-explaining context to AI agents
Cons
  • Requires Node 18+ and initial setup via command line
  • Screen recording permission required on macOS for full functionality
  • Deep memory (dreaming) features require sending data to external AI models, though audited

Best For

Coding agents (Claude, Cursor) automatically understanding what you were doing across appsPersonal memory assistant for recalling actions in any application with source and dateBuilding long-term user profiles for AI agents to personalize interactions

FAQ

What is Continuum?
Continuum is an open-source MCP server that captures your entire desktop context (screen OCR, audio transcription, typed content) and serves it as structured memory to AI agents, so they understand your workflow without manual re-explaining.
How does Continuum protect my privacy?
Continuum runs locally on-device. The memory store is read-only to agents, and an append-only audit ledger tracks exactly what data leaves the machine with a build-failing check. Optional encryption at rest is available. The tool does not claim nothing leaves; it shows you exactly what does.
Which AI agents work with Continuum?
Continuum works with any MCP-compatible agent, including Claude Desktop, Claude Code, Cursor, and Codex. You can also integrate it manually by adding the MCP server configuration.