Notely

AI & MCP Capabilities ​

Notely adopts a privacy-first, local-centric AI architecture. Rather than relying on intrusive in-editor generative sidebars, conversational interaction is driven by an embedded Model Context Protocol (MCP) server, while local background workers handle semantic search embeddings and knowledge graph relation discovery.

1. External AI Integration via Model Context Protocol (MCP) ​

Notely embeds a dual-transport Model Context Protocol (MCP) server running on port 3700 (http://127.0.0.1:3700/mcp and /sse). This allows external LLM agents and desktop AI clients (such as Google Antigravity, Claude Desktop, Cursor, and IDE extensions) to interact directly with your notes, tasks, and diagrams.

7 Enterprise Unified Tools ​

External AI assistants have access to self-contained, structured tools:

  1. search: Semantic hybrid search (FTS5 + BGE dense vector cosine similarity).
  2. read_note: Read note content, frontmatter, backlinks, and git revision history.
  3. edit_note: Create, overwrite, or surgical chunk-replace note text with atomic write safety.
  4. manage_tasks: Aggregate open/completed tasks, toggle status, and append checklist items.
  5. manage_diagrams: Read, create, and modify Mermaid, Draw.io, and Excalidraw diagrams.
  6. workspace_overview: Inspect file trees, broken link integrity, graph relations, and disk storage stats.
  7. git_control: Review stage status, commit diffs, commit history, and branches.

Standard MCP Prompts ​

Exposes ready-to-run interactive workflows to MCP clients:

  • summarize_note: Generate executive summaries from note contents.
  • plan_tasks: Extract action items and checklists from freeform notes.
  • explore_knowledge_graph: Traverse semantic relationships around specific concepts.
  • refactor_note: Clean up structure, formatting, and heading hierarchy.
  • daily_review: Synthesize recently modified notes and pending workspace tasks.

For full schema details and client configuration, see the Enterprise MCP Tools Reference and Developer MCP Guide.

2. Local-First Vector Embeddings ​

Notely indexes your workspace into {workspace}/.notes-app/ai-embeddings.db:

  • Local BGE Model: Uses BGE-small-en-v1.5 (~130MB) executed on-device via onnxruntime-node.
  • Hybrid Semantic Search: Combines full-text search (SQLite FTS5) with 384-dimensional dense vector embeddings for semantic similarity queries.
  • Background Utility Process: Document chunking and vector calculations run in a dedicated background worker to keep the editor UI fast and stutter-free.

3. Knowledge Graph Engine ​

Notely features an offline, 8-stage Knowledge Graph engine mapped into {workspace}/.notes-app/ai-graph.db:

  • Neural Zero-Shot Extraction: Powered by an offline gliner2-multi-v1-onnx neural model running locally.
  • Concept Deduplication & Alias Fusion: Automatically merges variations of the same concept (e.g. "SQLite DB" and "SQLite Database") using cosine vector similarity (>0.88 threshold).
  • Interactive Visualization: Explore entity clusters, backlinks, tags, and document references interactively from Workspace → Workspace Graph (Ctrl/Cmd + Shift + G).

4. MCP Tools Catalog & Diagnostics ​

Inspect and test AI operations inside the app:

  • MCP Tools Catalog (Ctrl/Cmd + Shift + M): View all registered MCP tools, test tool execution live with custom arguments, and inspect JSON responses.
  • MCP Diagnostics & Telemetry: Accessible from the Diagnostics panel, this dashboard displays live connection status, active client sessions, request rates, execution latency, and error logs.
  • AI Settings (Ctrl/Cmd + Shift + ,): Configure LLM provider API keys (Gemini, Groq, OpenAI), embedding engine mode, and Knowledge Graph confidence thresholds.