Shatter LLM Context Limits Without Sacrificing Privacy.
Intelligently split massive PDFs into model-optimized chunks for Gemini, Claude, and Copilot. Zero server uploads — your data stays in your browser while we prepare it for the frontier.
AI Document Sidekick
Built for what actually goes wrong
Massive PDFs exceeding the token or byte limits of AI models, causing ingestion failure or lost context.
Deterministic 3-pass processing (Analyze-Plan-Build) ensures every chunk fits perfectly within model-specific ceilings.
Security risks associated with uploading sensitive documents to third-party servers for processing.
Client-side splitting via pdf-lib and pdfjs ensures your bytes never leave the local browser environment.
What ships in the box
AI-Native Chunking Engine
Uses a deterministic character-to-token ratio (chars/4) to greedily group pages while respecting token, byte, and page-count ceilings for every major LLM provider.
Reasoning Continuity (Overlap)
Applies a sliding-window overlap (e.g., 2,000 tokens for Claude) to ensure downstream AI agents maintain reasoning context across chunk boundaries.
CLAUDE.md Harness Generation
Automatically generates a structured metadata map (CLAUDE.md) that tells AI agents exactly what each chunk contains, serving as a "table of contents" for long-context memory.
Built on enterprise-grade infrastructure
Every byte of every PDF stays inside the browser. The pdfjs analyser, pdf-lib splitter, and chunk packer all execute in the user's WebAssembly sandbox — the server only serves a static HTML+JS bundle. Built for legal, medical, and regulated teams who can't upload sensitive documents to third-party clouds.
Cloudflare global edge — innate DDoS protection, zero cold-starts.
Clerk (SOC 2 Type II, GDPR) for multi-tenant fleets; PBKDF2 for single-tenant apps.
Cloudflare & Resend (SOC 2 Type II) · Stripe (PCI-DSS Level 1).
From foundation to fleet scale
- Phase 1Shipped
Foundation — Client-Side Core
- 3-Pass PDF processing engine (Analyze, Plan, Build) in Vanilla JS.
- Support for Gemini, Claude, and Copilot constraint sets.
- JSZip-based bundle generation with zero-upload privacy guarantee.
- Phase 2In progress
Metadata & Discovery — MCP Integration
- Cloudflare Worker deployment for MCP (Model Context Protocol) metadata.
- JSON-RPC 2.0 endpoints for programmatic constraint discovery by external AI tools.
- Advanced configuration panel for custom overlap and ceiling overrides.
- Phase 3Next
Enterprise & Visuals — Enhanced Triage
- Visual chunk map preview with interactive page-range selection.
- Multi-PDF batch processing and cross-document relational mapping.
- Native integration with IDE-based AI agents (Cursor, Claude Desktop).
Ready to bring AI Document Sidekick into your stack?
Talk to our team about deployment, custom SLAs, and integration with your existing identity and observability stack.