SINGLE FILE · ANONYMOUS · ON-DEVICE
OWLAGENT
Decision-augmented auto-research & auto-plan agent.
Every question becomes a research run. OWL classifies the interrogative intent, expands it into sub-queries plus related terms, fans out across 24 free keyless public sources — Wikipedia, Wikidata, arXiv, OpenAlex, Europe PMC, Crossref, Semantic Scholar, GitHub, Hacker News, Reddit, Lobsters and more — scores each passage with a decision kernel calibrated on confidence, then answers in its own words. The result lands in a 384-dimension vector kernel persisted to IndexedDB. No server. No account. No upload. One HTML file.
Everything an agent needs, in one file.
OWL is not a chat wrapper. It is a complete research and planning stack — interrogative classifier, query expander, 24-source scraper, decision kernel, vector database, domain-aware planner, repo cloner, page reader and image-free export layer — compiled into a single HTML document that runs entirely in your tab.
Calibrated confidence
Every source, claim and plan step is scored on overlap, source weight, depth, trust, interrogative type-fit and freshness. Consensus, mixed evidence and contradictions are labelled across sources, and a source-graph bonus rewards diverse trusted origins.
CAD vector kernel
Every message becomes a 384-dimension embedding, cosine-searched in microseconds and persisted to IndexedDB. In-memory store capped at 2000 vectors with FIFO eviction; near-identical inserts are deduplicated by cosine similarity. Close the tab — the owl remembers.
24-source keyless fan-out
Weight-ordered fan-out with early exit once core full-text evidence is gathered, then a second multi-hop fan-out on the top entities it found. Rate-limited endpoints back off politely; every source soft-fails.
Auto-research loop
/auto <topic> picks angles by expected value
(novelty × confidence gap), chases entity gaps, and stops on a confidence plateau,
repeated errors or the cycle cap. One request for the whole session.
Domain-aware plans
/plan detects the domain — software, cooking,
mechanical, business, learning, travel, health, event, creative, general — asks up to
three clarifying questions, grounds each phase with a bounded research call, and emits
a domain-appropriate deliverable.
/git into IndexedDB
/git user/repo clones a repository into IndexedDB
with capacity estimation, warnings for oversized trees and graceful
partial-clone behaviour when storage fills up. Export any clone as a zip.
Weather · calc · digest
/weather uses Open-Meteo and never charges a
request. /calc is a local expression parser.
/digest fuses confident claims into one coherent
paragraph with a spread line. Always free, at any quota.
/page reads any URL
Direct fetch, then a public proxy, then a text reader proxy. OWL extracts the page, summarizes it locally with an extractive ranker, files a card and cites the exact link.
Folder · digest · exports
Every research turn, plan, auto session, git clone, page read and weather lookup
files a card. /digest fuses confident claims into one
paragraph. Export as markdown, folder zip or full archive.
Anonymous by architecture
Chat, kernel vectors, plans, folder cards and cloned repos never leave the browser. Only the query text touches public APIs. No account, no tracking, no server-side storage.
Living ASCII owl companion
One concise snowy owl rendered as a pure-ASCII character grid: it waits perched on its branch, soars over the evergreen forest on every request, and explores head-first past the tree line while working locally. Big googly eyes track your pointer; it blinks on its own rhythm. Every glyph glows with a hot white halo.
Live telemetry panel
/stats opens the live telemetry panel — session,
kernel, folder, auto, decision, storage and IndexedDB quota, all in one view. Also
reachable from the STATS button in the header or
ctrl+K.
How one question becomes a research run.
Six stages, all executed in your browser, all interruptible with
ESC.
Classify the interrogative intent
Before any network call, the turn is routed: how, why, when, who, where, which, what or plain search. Build / make / create / design questions fall to how. The type drives sub-query focus, source weighting and the answer's synthesis hint.
Expand into sub-queries
The query is rewritten into up to three focused variants — the original, a
type-focused variant such as x process steps mechanism,
and a Datamuse-related-terms variant.
Fan out across 24 free sources
Up to 28 jobs per turn, weight-ordered, with a hard 9-second research deadline, per-request timeouts, polite back-off and early exit once core full-text evidence is gathered. Then a second multi-hop pass on the top entities the first pass surfaced.
Score with the Decision Kernel
Each passage is scored 0–1 from query overlap (plural- and possessive-aware), source weight, text depth, source trust and interrogative type-fit, with a freshness bonus for recent items. A consensus detector labels strong consensus, mixed evidence or contradiction.
Synthesize locally
An extractive ranker picks the sentences, then a small local LaMini-Flan-T5-77M generator is consulted only when extractive confidence is low. If models can't load, OWL falls back to deterministic extractive synthesis — never to nothing.
Persist to the kernel and the folder
The answer, its top claims and every card are embedded and written to IndexedDB. On the next boot the kernel hydrates, the folder re-renders, and the cooldown counter resumes — offline recall works.
Every command. That's the whole manual.
Type /help in the app for this list, or just ask a
question — plain text is always treated as a research query.
<any text>Research a question — terse, decision-first answer with calibrated confidence and sources./plan <anything>Domain-aware plan — software, cooking, mechanical, business, learning, travel, health, event, creative or general./price <item>Real price match — FX via Frankfurter (ECB), crypto via CoinGecko, products via live DuckDuckGo scrape. Median, trimmed mean, range and links./auto <topic>Auto-research loop — one request total, stops on a confidence plateau./auto plan <thing>Growing plan — weaves confident findings into the next un-grounded phase./stopStop the auto session or interrupt the current turn./git <user/repo>Clone a repository into IndexedDB — export as zip. Capacity-aware, partial-clone safe. Free, single-use./page <url>Read any page: direct → proxy → reader. Files a page card. Free, single-use./weather <location>Free forecast via Open-Meteo geocoding + forecast. Never consumes a request./calc <expr>Local safe calculator: + - * / % ^ sqrt sin cos tan log ln abs round floor ceil pi e./digestFuse confident claims from the last research turn, auto session, recent cards and the kernel into one paragraph./sourcesList every free endpoint in the fan-out./statsSession + kernel telemetry (also ctrl+K)./folderBrowse cards, export md / zip / repo./export md [id] | folder | all | repo <id>Exports — always free./reset confirmWipe local data; the free-request window is per browser and survives a reset./helpPrint the full command list inside the app.Shortcuts
enter research · ↑ / ↓ prompt history · esc interrupt the running turn (or stop auto, or close a modal, or cancel a pending expansion) · ctrl+K telemetry.
Getting good results
Be specific — "history of the Antikythera mechanism" beats "old computers". Use /auto when you want breadth, /plan when you want order, and /page when a single result looks promising but thin.
Free to try. Cheap to own.
OWL runs on your hardware, so there is no inference bill to pass on. The demo is free with a 10-request cap per rolling 24-hour window. OWL PRO — unlimited research for $10/month — is under development and will open here when it ships. Donations are optional and do not unlock extra requests in the demo.
Try the entire pipeline — research, kernel, auto-agent, plans, repo cloner, page reader, exports — capped at ten counted requests per 24-hour window.
- 10 counted requests per rolling 24h
- 24 free keyless research sources
- Auto-research, plans and multi-hop
- Vector kernel + IndexedDB memory
- Repo clone into IndexedDB
- Free
/weather,/calc,/page,/git - Local models, loaded on demand
NO CARD · NO SIGN-UP · RESETS PER BROWSER
Unlimited research, unlimited auto-agent cycles, no request cap. Currently under development — the free demo shows the entire pipeline today.
- Unlimited research requests
- No 24-hour cooldown
- Unlimited auto-research cycles
- Full plan and repo execution
- All 24 research sources
- No telemetry, ever
- Cancel any time
UNDER DEVELOPMENT · PRICE MAY VARY
OWL is under development (webXOS 2026). Would you like to donate to help fund and speed up the process?
- Fund development and hosting
- Entirely optional
- Does not unlock extra requests
- Paid subscription stays planned at $10/mo
$1 · $5 · $10 OPTIONS IN-APP
Questions, answered plainly.
Is this a real product or a demo?
Where does my data go?
Do I need an API key or an account?
Which sources does OWL actually use?
How does the 24-hour cooldown work?
/weather, /calc,
/page, /git, /help, /folder,
/export, /digest, /sources, /stats,
/reset) always work at the limit.What is the difference between research, plan and auto?
/research runs a single
research turn and returns a terse decision-first answer. /plan
detects the domain, asks up to three clarifying questions when the goal is vague,
and compiles a phase-by-phase plan document — a repo tree only for software, a recipe
card for cooking, a BOM for mechanical builds. /auto loops the
research, picking angles by expected value and stopping on a confidence plateau. One
request for the whole auto session, charged only after the first successful cycle.
Auto supports research and plan only — /page, /git and
/weather are single-use and never loop.What does /git actually do?
/git user/repo fetches the repository metadata and
tree, checks how much room is left in IndexedDB, then clones the repository into
your browser — one file at a time, with a hard per-file size cap and a total
budget. If the repo is very large OWL will warn you, clone as much as fits, and file
a partial-repo card. You can export the clone as a zip at any time.What are the local models, and do they download?
Can it run offline?
/digest, /folder and
previously explored pages all work with no network. New research obviously needs a
connection.