clew. follow the thread.
Bulk-ingest the documents. Read every page. Ask questions that answer with page-level citations.
Read everything
Every page gets read
Drop up to 500 files at once: PDFs, images, spreadsheets, Word documents, email, HTML, plain text, or a whole zip. Each page climbs an escalating reader chain until it yields text, and a crashed engine falls through to the next one instead of killing the job.
Embedded text
digital pages are extracted directly
Local OCR
Tesseract or PP-OCRv5, on your CPU, offline
Vision model
hard pages escalate to a VLM, if you enable one
Four ways to search the corpus
Keyword, semantic, hybrid with reciprocal rank fusion, and raw regex. Scoped to the datasets you choose, reranked before you see them.
pages read in one test corpus
search modes
file format families
Answers that cite their pages
Chat with the corpus at three depths, up to durable research runs that survive a closed tab. The model can only cite pages it actually retrieved this turn; a broken citation fails visibly instead of pretending. Drag any citation straight onto the board.
Who authorized the transfer?
The transfer was approved on May 14 by the harbor office1, countersigned by R. Smith two days before the audit date 2.
Pin it to the wall
A spatial evidence board with nine node types: live document readers, entities, notes, frames, maps, timelines, drawings. Connections draw as red string, because that is what they are.
smith-deposition.pdf
p 41 · embedded reader
Transfer countersigned before the audit date.
Every name, place, date, and amount
Entities are extracted across six kinds, with dates and amounts normalized to real values. A five-round AI cleanup wizard merges duplicates and fixes types, and each investigation keeps its own independent entity set.
person
R. Smith
organization
Harbor Office
place
Pier 4, Gdansk
date
May 14, 1998
amount
$120,000
other
Case 44-B
Time as a surface
Dated evidence becomes a chronograph: datetime-scaled swimlanes with night shading, severity leader lines, and every event anchored to a cited page. Or let the model narrate a timeline, citations enforced.
For your agents
Your corpus, your agents
A built-in MCP server exposes six tools over streamable HTTP, with generated configs for nine clients and hashed, scoped, revocable keys. Search, read pages, look up entities, and fetch page images from any MCP-capable assistant.
{
"mcpServers": {
"clew": {
"type": "http",
"url": "https://your-clew.example/api/mcp",
"headers": { "Authorization": "Bearer <key>" }
}
}
}Self-hosted
your documents never have to leave your server.
Offline by default
Chat, embeddings, search, and OCR all run with zero API keys: a deterministic mock provider, local embeddings, and bundled OCR engines.
Cloud AI is opt-in
Point individual tasks at Anthropic, OpenAI, OpenRouter, or Ollama when you choose to. Until then, nothing is sent anywhere.
Everything in the box
Postgres and S3 storage ship with the install. One command to set up, verified backups before every upgrade, automatic rollback.
start pulling the thread.