Ziwei MCP (@lhk714/ziwei-mcp)

Deterministic Zi Wei Dou Shu (紫微斗数排盘) Model Context Protocol (MCP) server: the iztro star-placement engine wrapped in a real astronomical time layer, with every school convention exposed as an explicit parameter.

中文文档 (Chinese) | English

🌟 Overview

Every Zi Wei library in the ecosystem is ultimately iztro — it is the one real engine, and its star placement is correct. What has been missing is a server that feeds it the right time.

@lhk714/ziwei-mcp is that layer. It gives an LLM a correct chart for any birthplace on Earth, in any year from 1900 through 2100 (Beijing time), and names the school convention behind every switch instead of silently picking one.

Birth Wall Clock + IANA Timezone  (1990-06-15 20:00 America/Los_Angeles)
                          │
            ┌─────────────┴─────────────┐
            ▼                           ▼
   【Axis A: true UTC instant】   【Axis B: local solar time】
     decides the year ganzhi        longitude [+ Meeus EoT] − DST
    (正月初一 by default, or         decides lunar date + timeIndex
     the exact 立春 instant)          (solarTime: true/mean/off)
            │                           │
            ▼                           ▼
        year ganzhi  ───────────►  iztro star placement
     宫干 · 四化 · 年系星 · 大限      命宫 · 身宫 · 紫微系 · 昌曲空劫火铃

Why the two axes are separate

Zi Wei couples time to the chart in two different ways, and conflating them is the single most common source of wrong charts:

• The hour, day and month come from local solar time at the birthplace (Axis B, correction mode set by solarTime) — they decide the 命宫/身宫, the 紫微 and 天府 series, and 昌曲/空劫/火铃.

• The year ganzhi comes from the resolved birth instant in real physical time (Axis A) — it decides the 宫干 (五虎遁), all 四化, every year-based star, and the 大限 sequence. By default the boundary is 正月初一 (the mainstream 紫微斗数 convention — see yearDivide below); the true 立春 instant is available as an opt-in for the 八字/子平术-aligned boundary.

A birth two hours before 立春 in Los Angeles is after 立春 in Beijing. When yearDivide:'lichun' is in effect, only one of those is the year ganzhi, and getting it wrong invalidates the whole chart, not part of it.

Known engine bugs this server works around

Both were reproduced against iztro@2.6.0 and are covered by regression tests:

Two upstream calendar defects are also worked around: a leap month that does not exist is rejected rather than silently charted as the ordinary month, and 闰月三十 births (17 such days between 1900 and 2100, e.g. 2017-08-21) build correctly instead of throwing.

Related MCP server: AstroStellar MCP Server

🚀 Quickstart

bunx @lhk714/ziwei-mcp@latest

npx -y @lhk714/ziwei-mcp@latest

⚙️ MCP Client Configuration

{
  "mcpServers": {
    "ziwei": {
      "command": "npx",
      "args": ["-y", "@lhk714/ziwei-mcp@latest"]
    }
  }
}

@latest re-resolves from the registry on every client launch — that's deliberate, so you always get fixes — but it costs a network round-trip at startup and fails hard offline. For an offline or latency-sensitive setup, pin an exact version instead, e.g. @lhk714/ziwei-mcp@0.2.0.

🛠️ Tools

1. calculate_ziwei

Requires gender, one of solarDate/lunarDate, one of clockTime/shichen, and a location (place, or longitude + timezone).

Location — place (English city name), or longitude + timezone (IANA name). A timezone is never inferred from longitude.

Date & time — solarDate or lunarDate (+ lunarDateFrame: local | beijing); clockTime or shichen; dstFold (0/1) to disambiguate a DST fall-back hour.

School conventions — all optional, all reported back in diagnostics.convention:

mutagens and brightness are passthroughs to iztro's own tables, so the default is whichever table the pinned iztro version ships (diagnostics.engineInfo.iztro). When either is supplied, diagnostics.convention.brightnessOverrides / .mutagenOverrides name the stars and stems that came from the caller — an overridden 庙 and iztro's own 得 are the same shape in the output, and the 运限 path re-applies the same tables, so both tools disclose it.

Every convention default here was audited against the ecosystem (iztro's own factory defaults, 测测, ziwei.pub) as of 0.2.0. dayDivide and fixLeap previously diverged from iztro's own factory defaults by accident, not by deliberate school choice — that has been corrected. yearDivide/horoscopeDivide defaulting to lunar_new_year is the project owner's doctrinal ruling: 立春 belongs to 八字/子平术, not to this system's own star-placement apparatus. yearDivide:'lichun' remains fully supported — the Z1 machinery behind it is unchanged and still, as far as this project has found, the only correct 立春 implementation in the ecosystem.

algorithm: 'zhongzhou''s behavior has been exhaustively mapped against iztro, but only compared to secondhand online sources for Zhongzhou-school (王亭之) doctrine, not the original texts. One known contradiction: it does not change 四化 (庚/壬 stay on the textbook table, not the documented Zhongzhou one) — use config.mutagens if you need Zhongzhou 四化. astroType:'earth'/'human' combined with algorithm:'default' is rejected outright (not just unverified): it would return a chart whose 命主 contradicts its own 命宫.

Not supported: unknown birth time. 命宫, 身宫, 文昌/文曲, 火星/铃星 and 地空/地劫 all depend on the hour. Without it there is no chart, so the request is rejected rather than answered with a partial one. shichen is supported; when true-solar correction pushes a spoken shichen across a boundary, the response carries shichenAmbiguity with the candidate timeIndex values instead of a confidently wrong single chart.

Output — trimmed to keep an LLM's context usable: the 12 palaces with their stems, branches, major/minor/adjective stars, brightness, 四化 and 大限 range, plus 命宫/身宫, 命主/身主, 五行局, the lunar date, and a diagnostics block recording the exact instant, both axes, the longitude and equation-of-time corrections, the conventions applied, and any warnings.

2. calculate_ziwei_horoscope

运限 — the moving chart: 大限 (decade), 小限 (minor year), 流年/流月/流日/流时 (year/month/day/hour), each with its own 四化 and 运曜. Takes the same birth contract as calculate_ziwei, plus a target (solar date + clock time, resolved through the same time layer — solar time correction mode, IANA, DST). Omit target for "now".

It is a separate tool on purpose: folding six scopes × twelve palaces of 运曜 into the natal response would blow up an LLM's context for callers who only wanted the chart.

iztro's 运限 arithmetic is sound — an independent implementation of the classical rules agrees with it. The CI-gated suite checks 200 seeded charts × 4 targets on every run; a one-time full manual sweep (bun run tests/horoscope-parity-reference.mjs 800 10) went further, agreeing across 303,582 assertions, zero mismatches. Its interface needed wrapping, and this tool does it:

• The year-ganzhi bypass that keeps the natal chart correct silently poisons every age-derived scope, because 虚岁 is target lunar year − fed lunar year + 1. Compensated per scope — 流月/流日/流时 always come from the true target, since 流日 is JDN-based and not 60-year periodic.

• horoscopeDivide is locked: under iztro's 'exact', 流年 divides at 立春 while 虚岁/大限/小限 divide at 正月初一, so one response contradicts itself six days a year.

• iztro's config is global and horoscope() reads it lazily, so one caller's school override would otherwise rewrite later callers' charts.

• ageDivide: 'birthday' is rejected here: it flips on the 1st of the month after the birth month and ignores the birth day, so honouring it would silently mean something other than what it says.

• Targets before the birth are rejected (iztro returned index: -1 and untranslated i18n keys, silently), and a late-Zi target is normalized (dayDivide does not affect horoscope() at all).

Under yearDivide:'lichun', age reckoning uses the 立春-designated birth year against a 正月初一 target axis. That asymmetry is a deliberate, documented convention choice, not an accident — the diagnostics report it, and the tests pin it. (Under the default yearDivide:'lunar_new_year', the birth side is simply the true lunar year — there is no asymmetry to reckon with.)

3. lookup_location

Resolves an English city name to longitude, latitude and IANA timezone across 7,329 cities in 227 countries.

Ambiguous names are refused, never guessed. "Los Angeles" exists in both the US and Chile; "Columbus" is two cities 2° of longitude apart that happen to share a timezone — 8 minutes of true solar time, enough to cross a 時辰 boundary. Sharing a timezone is not being the same place. Entries at genuinely identical coordinates (Kansas City MO and KS) still resolve: recognising that two records describe one location is a fact about the data, not a guess about intent.

A refusal returns a structured payload the calling agent can act on without parsing prose:

{
  "code": "ambiguous_place",
  "message": "…",
  "matched": 4,          // true hit count, so a capped list never reads as exhaustive
  "candidates": [ { "name": "San Jose", "province": "California", "country": "US",
                    "latitude": 37.3, "longitude": -121.85, "timezone": "America/Los_Angeles" } ]
}

Candidates carry identifying fields only. Population is deliberately absent: it is a likelihood prior, not something anyone recognises their birthplace by, and publishing it would move the guess this server refuses to make into the agent's prompt.

🧭 Four Pillars

This server does not calculate Bazi. Use @lhk714/bazi-mcp — it shares this time layer, so the two agree by construction.

🧪 Verification

bun test — every test below passing, including:

• An independent implementation of the classical star-placement rules (安星诀) from the source texts, checked against iztro across hundreds of seeded random charts and ~18,000 assertions. This is what pins the engine: an upstream change to any star's placement fails the suite.

• 立春-boundary scans asserting the year ganzhi flips exactly once, at the true instant — the regression guard for Z1.

• Differential batteries over 20 regular and 20 pathological timezones (date-line crossings, 45-minute offsets, historical upheavals), DST gaps and folds, leap months, and both lunar date frames.

• The full input-validation matrix, and real stdio end-to-end tests against the built binary.

External comparison against ziwei.pub (iztro's own demo) is manual and not part of the automated suite.

📜 License

MIT. Star placement by iztro (MIT).