Reckons.AI

🌐 Live app: reckons.ai

A personal knowledge graph that turns documents, web pages, and notes into a structured graph of facts — stored as a standard Turtle (.ttl) file on your device.

You own the file. You review and confirm every fact. Nothing leaves your device unless you choose to share it.

🚧 Alpha — feature-complete for personal use and actively developed. Expect rough edges; please file issues.

What it does

note / url / doc / .ttl / calendar / extension
        |
        v
   AI extraction
        |
        v
   Review & confirm  <-- only what you approve enters your KB
        |
        v
   3D knowledge graph   Reckoning (decision support)   Share .ttl

1. Ingest — paste a note, URL, document, calendar, or an existing .ttl file from someone else
2. Review — confirm or reject each proposed triple; refine labels before accepting
3. Explore — 3D force-directed graph with hub emphasis, layout modes, filter chips, and 2D fallback
4. Reckon — describe a situation and a target; the AI proposes options grounded in your KB
5. Compare — import a .ttl file or analyze a page to see what's new, conflicting, or reinforcing
6. Share — export .ttl and send it to collaborators; they see the diff on import

Key features

• Multi-KB management — create, switch, rename, and delete independent knowledge bases
• KB Leap — cross-reference entities between KBs; click to jump between them
• Browser extension — compare any webpage against your KB, accumulate research sessions across tabs, batch ingest
• MCP server — expose your KB to Claude Desktop, Cursor, and other MCP-compatible AI agents
• Predicate Manager — view, rename, and merge predicates across your KB
• Content safety — ethics preamble in all LLM prompts, content classifier, export advisory
• Passage grounding — verbatim source excerpts attached to extracted triples
• Diff summaries — LLM-generated 3-part summaries (new/reinforcing/conflicting)
• Whisper STT — local speech-to-text via transformers.js (no cloud required)
• Kokoro TTS — local text-to-speech for story walkthroughs
• Per-task LLM backends — use different providers for ingest, chat, analysis, and diff summary
• Model cache management — inspect, sideload, and purge locally cached WASM models
• Source trust system — sources earn trust through consistent accuracy; trusted sources auto-confirm
• History mode — time-travel through your KB with a timeline scrubber
• Cross-KB alignment — align entities across knowledge bases with embedding similarity and IRI remapping
• n8n cloud sync — private cloud sync via self-hosted n8n VPS; upload, download, and monitor KBs
• Source monitoring — watch URLs for changes, detect diffs, queue pending notes automatically
• GitHub repo ingest — ingest repository structure and code as knowledge with delta compare
• Entity normalisation — embedding-based IRI rewriting prevents duplicate entities at ingest time
• Self-dogfooding MCP workspace — the product tracks its own roadmap and status via its own MCP server

Quick start

cp .env.example .env   # add at least one AI backend key (or leave blank for WASM)
pnpm install
pnpm dev               # http://localhost:5173

No AI key required — the local WASM backend works out of the box (slower, fully offline).

For Docker:

docker compose up      # http://localhost:5173

AI backends

Full setup details in SETUP.md and docs/GUIDE.md.

Tech stack

• SvelteKit 2 + Svelte 5 (runes) — frontend framework
• Threlte 8 / Three.js — 3D force-directed knowledge graph
• Dexie — IndexedDB persistence (all data stays in-browser)
• N3.js — W3C RDF/Turtle parsing and serialization
• @huggingface/transformers — local WASM LLM and embedding inference
• Playwright — end-to-end test suite

Data model

Every fact is a Statement — an RDF triple with provenance:

{
  s: { kind: 'iri', value: 'urn:kbase:person/alice' },
  p: { kind: 'iri', value: 'urn:kbase:predicate/organized' },
  o: { kind: 'iri', value: 'urn:kbase:event/float-trip' },
  g: { kind: 'iri', value: 'urn:kbase:source/<uuid>' },  // provenance
  sourceId: '<uuid>',
  confidence: 0.95,
  status: 'confirmed',   // pending | confirmed | refined | rejected | superseded
  excerpt: 'Alice organized the float trip last summer.',  // verbatim source sentence
}

Exported as standard Turtle (.ttl) — readable by any RDF tool.

Browser extension

The extension adds a side panel with three tabs:

• Compare — analyze the current page against your KB with at-a-glance proportional bar
• Session — accumulate findings across multiple pages with aggregate summaries and batch ingest
• Ingest — send extracted triples to Reckons.AI

Supports Chrome, Edge, Brave, Firefox desktop, and Firefox for Android.

See SETUP.md for installation instructions.

MCP server

The standalone MCP server (mcp-server/) exposes 20 tools to AI agents:

Self-dogfooding workspace

Reckons.AI uses its own MCP server to track product state. Six internal KBs (Roadmap, Production, Features, Architecture, Testing, Codebase) are symlinked from static/*.ttl into mcp-workspace/kbs/. Claude Code queries these KBs before planning new work or modifying architecture. Edit a TTL file → the MCP server auto-reloads → the next AI session sees the change.

bash scripts/setup-mcp-workspace.sh   # one-time setup after cloning

Documentation

Documentation KB

The app ships with a built-in documentation graph (starter-guide.ttl) containing:

• Core philosophy and getting-started steps
• KB Leap nodes linking to 7 deep-dive sub-graphs (auto-imported on first click):
  • Triples & RDF standards
  • Language models & RAG
  • Real-world use cases
  • All features
  • Integrations & technology
  • Tips & security
  • Timeline & RDF ecosystem
• A guided story walkthrough narrated by Shelly

Context compression

Knowledge graphs are dense by nature. Reckons.AI compresses what you know into structured RDF triples — retaining semantic meaning while reducing the tokens an AI needs to understand your situation. A page of prose becomes a handful of triples. Same meaning, fraction of the tokens. Feed your compressed KB directly to AI agents via MCP.

Enterprise roadmap

Structure organisational knowledge around People · Policy · Procedure — the three dimensions that matter. RBAC, bring-your-own auth (SSO/LDAP/OIDC), file-based .ttl delivery, and self-hosted deployment via n8n. See docs/ENTERPRISE.md.

Things that are deliberately absent

• No backend server — all state in IndexedDB; Turtle export for backup. Optional n8n cloud sync is self-hosted.
• No accounts — your KB is yours alone, on this device. Enterprise RBAC is an opt-in layer.
• No analytics, no tracking, no remote logging
• URL ingestion proxies through r.jina.ai/<url> for clean-text extraction; use the note or document tab to avoid that hop entirely