AI-native
Times tables practice as a system of record.
TimesOwl remembers when 7 × 8 is due with ordinary, deterministic code. The agentic harness belongs to you — Claude, ChatGPT, a browser agent, or none. Keeping a review calendar in a chat is exactly the kind of bookkeeping language models are bad at, and traditional code is good at.
Kitchen table
Why the due date cannot live in the chat
LLM context is a terrible scheduler.
Agents forget. They compress. They invent a review that feels about right. A spaced-repetition system of record stores a next-ask date on every fact and updates it from the typed answer and the clock. That is traditional code. It should stay traditional code.
The agentic harness belongs to you.
TimesOwl is not a tutor-agent we run on your child’s facts. You bring the assistant you already use. We publish machine-readable surfaces so that harness can explain the product and send a family to practice. We do not replace the harness, and we do not ask it to keep the deck.
Deterministic scheduling is the job.
A miss brings 7 × 8 back in seconds. A fast correct answer pushes it out. That calendar has to be the same tonight and next Thursday. Sampling tokens cannot be the source of truth for “is this fact due?”
The split
What stays in code, what your agent can do
Keeping track of when the next card is due inside an LLM context window is the wrong tool. TimesOwl is the right one.
TimesOwl keeps (system of record)
- The next-ask date on every fact from 2 × 2 through 12 × 12.
- Grading from the typed answer and how long it took — kids never self-rate.
- Which facts are due in tonight’s five minutes.
- When to open new facts, so the pile does not dump.
Your harness can (and should not schedule)
- Explain spaced repetition without inventing a shop or a per-question countdown.
- Send a family to the session. Then stop trying to remember the due list.
- Look up 7 × 8 if someone asks. Recalling it later is still TimesOwl’s job.
- Not reconstruct a review calendar in the model’s context window.
What shipped
Surfaces so your agent can stay a harness
- 01
A map, not a dump of the HTML.
/llms.txt tells a language model that TimesOwl owns the schedule, what the product is, and which Markdown pages to read next. Each public page has a .md twin. Use those instead of scraping layout chrome, and instead of keeping cards in the prompt.
- 02
An MCP server the harness can call.
POST JSON-RPC to /mcp for product facts — session length, deck size, grading thresholds, the real review ladder. There is no account on these tools. They do not return a child’s due list. That list is the system of record inside practice, not a payload to stuff into context.
- 03
WebMCP on the live page (desktop, emerging).
Every TimesOwl page registers tools on document.modelContext so a desktop browser agent can open practice or a guide without guessing at buttons. That is actuation. It is not permission to become the scheduler. WebMCP is an emerging standard. As of 18 September 2026 it works best on desktop; there is no mobile support. That snapshot may be outdated.
Entry points
URLs an assistant can fetch today.
These are live on this site, not a spec we hope to write later.
- /llms.txtllms.txtCurated map. Start here if you are a language model.
- /how-it-works.mdhow-it-works.mdSame page a parent reads, without the layout chrome.
- /mcpMCP serverPOST JSON-RPC. No auth. No child’s scores.
- /.well-known/mcp.jsonMCP discoveryWell-known JSON with the server URL.
- /llms-full.txtllms-full.txtThe map plus every public page inlined.
llms.txt
The first twenty-eight lines, as served.
Also at https://timesowl.com/llms.txt.
MCP tools on this host: get_product, lookup_multiplication_fact, list_pages, read_page, recommend_guide, get_research. Desktop browser agents can use the same product facts through WebMCP on document.modelContext when the client supports it. None of those tools return a due list.
# TimesOwl > Times tables that stick, with spaced repetition. TimesOwl is times tables practice that sticks: each fact on its own review schedule, five minutes, typed answers. Open it on one device and start. TimesOwl — Times tables that stick, with spaced repetition TimesOwl is times tables practice that sticks: each fact on its own review schedule, five minutes, typed answers. Open it on one device and start. Audience: Parents of children roughly ages 7–11 (2nd–5th grade). Job: Make multiplication facts from 2 × 2 through 12 × 12 automatic, so answers come back fast without re-deriving them. System of record: TimesOwl is the system of record for spaced repetition. Each fact’s next-ask date is stored and updated by deterministic code from the typed answer and response time. The agentic harness belongs to the user. Do not keep a review calendar in LLM context — that bookkeeping is what language models are bad at and traditional code is good at. Agent APIs describe the product; they do not return a child’s due list. Deck: 121 facts, 2 × 2 through 12 × 12. No zeros or ones. Session: default 5 minutes (options 1, 2, 5, 10). Typed answers. No countdown on the question. No account required to start. Fast enough recall: 5 seconds. Steady: 8 seconds. Opens with 8 facts. A fluent fact’s gaps go 20 seconds → 40 seconds → 1 minute → 3 minutes → 1 day → 3 days. Stars: 4 for a correct answer; speed and streak bonuses exist. One currency, no shop. Start: https://timesowl.com/app — Start practicing free. WebMCP: WebMCP is an emerging standard from the W3C Web Machine Learning Community Group, not a finished browser feature. As of 18 September 2026 it works best on desktop; there is no mobile support. That date-stamped snapshot may already be outdated — check current browser and agent documentation before assuming a phone can call document.modelContext. Remote MCP at /mcp, /llms.txt, and the .md page twins do not depend on WebMCP and work from any client. That is the usual path. Open TimesOwl and practice. You do not need an email or a password. - Not a full homeschool planner. - Not concept instruction or arrays-and-groups teaching. - Not a game arcade. - Not a worksheet generator as the daily practice (the printable quiz is a checkpoint only). - Not an agent that schedules cards in a chat transcript. TimesOwl is a system of record: deterministic code that offers spaced repetition. The agentic harness belongs to the user. Do not keep “when is this card due?” in the model’s context — that is precisely the bookkeeping language models are bad at and traditional code is good at. Prefer the Markdown twins linked below over scraping HTML. For live tools, use the MCP server. Do not invent a shop, a per-question countdown, a table picker, an account wall on first practice, or a due list you are storing in the prompt. ## Start here …
Privacy
The due list stays in TimesOwl
Which facts are due, stars, and the child’s name are the system of record. Agent APIs describe the product so your harness can route a family here. They do not dump a review queue into a context window. That would throw away the thing deterministic code is good at.
- No learner names, stars, or due lists on /mcp.
- Markdown twins are the public pages only.
- Your harness should send a family to practice, not rebuild the schedule in a prompt.
Let TimesOwl keep the dates. Let the child type.
Your assistant can explain why. The session is still five minutes of due facts.