GitHub ↗leg 01
Upload
Bring the raw terrain: docs, notes, transcripts, repositories. Files wait in quarantine until they're scanned and promoted.
Knowledge bundles for humans and agents
Put everything you know on one map.
OpenKnoll gathers your scattered docs, decisions, and notes into a canonical, citable knowledge bundle — then answers questions about it with the evidence to prove it.
Sign in with your Google account. Nothing is uploaded until you say so.
The route
Four legs from scattered files to cited answers.
leg 02
Build
The survey pass turns your sources into an OKF bundle — canonical, versioned, and checksummed, so every claim has a fixed address.
leg 03
Explore
Walk the bundle with deterministic tools: browse its structure, follow references, and see exactly what made it in.
leg 04
Ask
Ask in plain language. Every answer arrives with citations that point back to the exact passage supporting it.
Evidence-backed
Answers you can check, not just believe.
you asked · why does the deploy gate wait for the scan?
Uploads land in quarantine and are only promoted after the malware scan and size checks pass1; the deploy gate reads the promoted revision, never the quarantine bucket2.
1ops-handbook.md § intake — “nothing serves from quarantine”
2pipeline-design.md § revisions — “gates resolve the promoted revision id”
you asked · what is the Zanzibar revenue forecast?
Insufficient evidence in this bundle to answer. No concept mentions: forecast, revenue, zanzibar. Adding a source about revenue planning would resolve the gap.
∅no citations issued — when the evidence isn’t there, it says so
Why survey at all
A folder of docs isn’t a map.
Raw folders and vector indexes fail the same way: the agent answers confidently from stale, unattributed, or missing knowledge — and nobody can tell.
No provenance
An answer synthesized from a folder or a top-k of chunks can’t say which source, which version, or whether it was ever reviewed.
No abstention
Similarity search always retrieves something, so the model rarely says “that isn’t in the knowledge base” — it improvises.
No quality gate
Duplicated, contradictory, or abandoned documents are served with the same confidence as the canonical ones.
No reproducibility
Ask again next week — different chunks, different answer, and no way to diff what changed in between.
OpenKnoll’s answer is a compile step with a contract on the output: defects surface at build time, provenance is pinned to content hashes and commit SHAs, and every revision is immutable and byte-for-byte reproducible.
What a container image is to “works on my machine”, an OpenKnoll bundle is to “the agent read our docs”.
Take it anywhere
The same map, in any agent’s hands.
A bundle isn’t locked to this app. Serve it over MCP and any agent explores it through seven bounded, read-only tools — it answers only from what it retrieved, cites bundle paths, and can’t follow a link off disk or reach the network.
overviewlistsearchpeekreadlinkshistory
Choose your route
Two routes up the hill.
Same format, same seven tools, same evidence rules — what differs is where it runs. Start on either route and switch anytime: the bundle is the contract.
OpenKnoll on the web
Nothing to install. Upload in the browser — files are quarantined and scanned before anything serves — build on the hosted pipeline, then explore and ask with citations. Share knowledge bundles with other users, and let claude.ai or any remote agent read yours through the hosted MCP gateway.
The CLI, on your machine
Apache-2.0 and fully useful offline — no account, no cloud service. Private by default and scriptable end to end: lint bundles in CI, prove rebuilds byte for byte, version everything in Git.
brew install openknoll/tap/oknoll oknoll init handbook && cd handbook oknoll add ./docs oknoll build oknoll ask "How are credentials stored?" oknoll serve --mcp
Platform API keys
Create, consume, and expose knowledge bundles on the cloud from your own code — build bundles from your pipelines and serve them to your agents with a key.
The datum
The format is a standard we don’t own.
Every survey is drawn against a datum — the reference frame all maps agree on. Ours is OKF, the Open Knowledge Format: an openly published bundle spec. OpenKnoll writes conformant bundles, and any OKF reader can use them.
Open Knowledge Format
Plain-Markdown concepts and references, a checksummed manifest, and defined cross-link rules — the whole format fits in one document, published upstream.
What OpenKnoll adds
The spec defines what a valid bundle is. OpenKnoll adds both the compiler and the runtime: builds that surface defects and pin immutable, byte-reproducible revisions — then a place to explore the result, ask with evidence, share it, and connect agents to it over MCP, hosted or on your machine. The next oknoll-python release grows the runtime further.
Either way, the bundle is yours: plain Markdown plus a checksummed manifest. Version it in Git, lint it in CI, pack it, hand it to any MCP client — a valid bundle stays useful without an OpenKnoll account.