Models guess your chart.Lumin computescomputesfixesexplainscomputes it.

No. 132 of 237 tools are KP; 61 Parashari, 20 KP-extended, 6 Jaimini, 5 Tajik, 3 Lumin extension, 10 composites that run several traditions in one call, and every response says which. More traditions are in the works, and so is the research on auditing one question across several of them at once.

Anyone whose question starts from a birth time or a moment: a chart, a match, a date to choose, a question asked right now. 237 tools cover that ground. People use it every day for their own questions. Product teams build it into their app, a date picker that knows what the date is for, a compatibility badge, a candidate fit score. And builders give it to agents, their own personal agent or a helper for an agent they already run, because an agent calls one tool, reads the result, and picks the next. Each result is more than a number. It says what was computed, why, and where the rule comes from. 9 open source apps and 13 situation pages show it in use, and the docs cover the rest.

An astrology app gives you the conclusion. Lumin gives you the conclusion and the working under it: every value it rests on, the rule that produced each one, what that rule ruled out and why, and the printed page the rule comes from. When two traditions disagree, you get both answers, each labelled. And anyone with the same birth details can run the calculation again and check it.

A model has read about astrology. It has never computed a cusp. Ask it twice and you get two charts, both stated with the same confidence. Lumin computes the chart, so the model reasons from the sky instead of from a guess. Ask again tomorrow and the numbers are the same.

Two things. It chooses which question to ask next, and it writes the final words. Every number in between comes out of a function. The server sets the order of the tools and the minimum number of calls, so the model cannot stop at the first answer that sounds right.

A library tells you where the planets are. That is the small part, 27 astronomy files against 237 rule modules. The rules, which one a position selects, whether it grants or denies, and which page states it, exist in no library. We still use the libraries, as a check rather than a source, and publish how far our numbers sit from theirs.

Yes, on purpose. Ask one matter across every tradition, and a tradition that has no procedure for a verdict says so and returns its findings, instead of inventing one. A rule nobody has traced to a source is marked untraced, never dressed up as certain.

No, and it does not claim to. Lumin checks that the calculation is right and that each rule is applied the way the books state it. Whether the tradition itself is true is a separate question, and not one we answer.

An investigation, not a lookup. The method runs in seven phases. Foundation, promise, timing, transit, verification, advanced techniques, synthesis. A full life-area reading is at least 20 tool calls and usually 25 to 40. One focused question is at least 12, usually 15 to 20. A factual lookup is at least 5, usually 5 to 8. No verdict is stated before the promise phase returns, and no date before both the dasha and the transit tools return.

Because nothing is stored and nothing is guessed. The server keeps no session and no memory. Every call recomputes the chart from the birth data you send, with the same ephemeris and the same rules. A paged result is identical on Tuesday and on Friday, page for page.

Every response carries what decided the verdict and what did not. A promise verdict names the gate it passed, reports the star lord route and the planet's motion without letting either gate the verdict, and cites the book and page the rule came from. A score states its scale and says it is not a probability. You can check any line against the source.

No. A result that does not fit is paged, never truncated. Every page carries a note naming the exact next call, so nothing is dropped or summarised. The one compaction left is on the full chart, and it names the tool that returns each field it leaves out.

Against the sky and against the books. The planets sit within 0.166 arcseconds of JPL Horizons over 14 instants from 1900 to 2050, and 6 observed sky events pin them and the Moon a second way. The Placidus cusps are checked against their own defining equation, to 0.16 arcseconds, and against Swiss Ephemeris, to 0.05. The KP tables and the dasha arithmetic are pinned to printed pages. 178 findings are counted, fixed and open alike, and the benchmarks page lists what is still open.

Add Lumin as a connector in Claude and sign in with your email address and a six digit code. See the setup walkthrough. In Claude Code, install the plugin instead. It adds the tools, the method, and a command per topic. Read the plugin guide.

Create an API key on the API keys page of the app and send it as a Bearer token, it starts with mcp_, to the /mcp endpoint. Keys are named and revocable. The 9 example apps show the wiring. Start building.

Yes. Lumin is a standard MCP server over Streamable HTTP. Any client that speaks the protocol can call it with an API key or through the sign in flow. The tool list and the reading plan arrive as tool results, so a client that shortens server instructions still gets the method.

300 tool calls a month, no card. A reading is 25 to 40 calls, so that is 7 to 12 readings every month. The allowance refreshes on the 1st.

From $1 for 400 calls. Packs are prepaid, drawn only after the free allowance is spent, and never expire. See pricing.

No. Only a call that returns a result counts against the allowance or a pack. A paid call that fails is refunded.

Request a top-up in the app. We arrange payment with you by email, or on WhatsApp if you give us a number, and card payments go through a payment link. We never see or store your card.

No. In the app, answers are written on your own model key, so that cost sits on your key and never passes through us. Lumin counts only the tool calls, the same way everywhere.

Through the MCP server, they are used to compute the result and then discarded. We do not store the details a call carries or the result it returns, and they are not written to our logs. The Claude Code plugin keeps your birth details in a file on your own computer, not with us.

The charts you save and your conversations, so you can come back to them, until you delete them. Usage records keep the tool name, the time and whether it worked, never a birth detail, a question or an answer.

No. No ads, no marketing profiles, no selling, and no training on your charts, questions or answers. What we measure in the app stays in our own database and never includes anything you type.

It is stored encrypted, decrypted only to answer your own questions, and shown back to you by its last four characters. You can remove it in Settings at any time. API keys are stored as one-way hashes, so even we cannot read them.

Delete your account in Settings. That removes your charts, conversations, model key and sessions at once. API keys are separate, so delete those first. To get a copy of your data, or to ask what remains, write to us and a person replies within 30 days. The privacy policy lists every record and how long it is kept.