Wenbu · 问卜

Wenbu · 问卜

免费排盘、抽牌与查阅资料，把结果变成可以继续思考的问题。

Free charts, tarot and source-based learning for a question you want to explore.

中文网站 · English website · 知识手册 · Documentation

Open source / 开源接入

MCP quickstart · CLI · Agent Skill · Releases · Contribute

Official MCP Registry record · app.wenbu/mcp. Domain ownership is verified; this directory record does not imply endorsement or an installation. 官方目录登记已完成，可核对远程地址、版本与源码。

Connect remote MCP at https://wenbu.app/mcp, use the dependency-free CLI, or install the Skill from skills/wenbu/. Read the integration guide to choose a connection and run a reproducible example. The full application and integrations are MIT licensed. This is a public repository, not an anonymous-data or feedback export.

第一次用，从这里开始

你想做什么	建议入口	需要准备什么
每天留一点新的视角	日签	不用填写资料；免费揭牌、记录和回看，不消耗对话回合
还不知道怎么问	Agent 对话	直接提问、选择一个开场，或先看不耗回合的完整示例
围绕眼前一件事思考	塔罗 / 易经	一个具体问题，不需要出生资料
看出生资料对应的传统命盘	八字 / 紫微	公历生日与当地时间；八字还需时区，紫微必须知道时刻
弄懂术语与计算规则	知识手册	可以从入门内容读起，再核对结果里的计算约定

传统符号适合学习和自我反思。历法计算能否复核，与解释能否预测现实是两件事；问卜不把排盘结果当作人生结论。

有哪些功能

• 日签：每天一张固定正位牌、一个小行动和一个回看问题，配原创牌面与克制揭签动效。无需资料或模型回合，可选云端保存、笔记和带原牌继续对话。Daily card in English.

• 八字：四柱、可见五行计数、未知时辰处理、午夜或子初换日选项，以及可选的近似真太阳时修正。五行数量不等于旺衰或喜用神。

• 易经：三钱法起卦，保留自下而上的六爻、动爻、本卦与之卦。

• 塔罗：完整 78 张牌，不放回抽取 1 张或 3 张，可选逆位；牌面为原创 AI 插画。

• 紫微：十二宫与星曜，公开 iztro 的排盘约定；目前宫位和星曜名称保留中文。

• Agent：调用实际工具、读取站内文章与精选外部资料、生成带来源的札记，并继续修改已有结果。研究范围限于资料目录。

• 保存与接入：可选邮箱账号与云端记录、浏览器试用手记、会话导出、MCP、CLI、Skill 与 REST API。

免费与资料处理

计算、抽取、学习内容、手记和 MCP 无需注册或付费。浏览器试用和登录账号每天最多 5 次 AI 解读、12 个 Agent 回合，上海时间零点重置；登录会归并当天试用用量，并受额外的网络防滥用和全站预算限制。外部匿名 API 仍按网络限额计数。AI 不可用时，独立工具仍可使用。详见免费额度。

未登录时，会话与手记暂存在当前浏览器。邮箱验证码登录后可开启云端记录、选择导入旧记录，并跨设备回看；不要求注册才能查看完整结果。计算请求会发送至问卜的 Cloudflare 服务；使用 AI 时，消息和所选上下文会经该服务发送给 DeepSeek。导出文件可能包含私人资料，下载不等于自动分享，也不等于匿名化。详见隐私说明。

Start with a question

For a short daily practice, try the free daily card: one stable upright card for the local date, an original action prompt and a question for later. No birth details, account or Agent turn required. Keep a note, optionally save to an account, or explicitly carry the same card into a conversation. It is a reflective practice, not a personal horoscope.

Use Wenbu Agent if you would like help framing a question. Ask directly, select a starter that sends the displayed message, or preview a complete example without using a turn. Selecting a starter preserves any draft you are writing.

For a question about the present, tarot and the I Ching need no birth details. BaZi and Zi Wei arrange traditional charts from birth information. The handbook explains how to begin and what the results mean.

The tools calculate or draw the symbols; AI can help discuss them. These are cultural and reflective practices, not scientifically established ways to predict a person's future. Visible element counts, for example, do not establish element strength or favorable elements.

Run locally / 本地运行

Requires Node.js 22.12 or later and npm. From this repository:

npm ci
npm run verify
npm run preview

Open localhost:8787. verify checks types, runs tests, builds the site and audits links and card assets. preview serves that build and the Worker API. Rebuild after changing static pages. npm run dev provides the faster Astro UI development server but does not supply Worker API routes.

AI is optional. To enable it locally, put server-only DEEPSEEK_API_KEY and QUOTA_SALT in an ignored .dev.vars file. No key belongs in browser code. See operations for setup, deployment and rollback.

Connect your own agent / 接入自己的 Agent

Use Streamable HTTP at https://wenbu.app/mcp. It exposes seven tools:

Tool	Purpose
calculate_bazi	Four Pillars with time-zone and day-boundary conventions
cast_iching	Random three-coin cast or six supplied lines, bottom to top
draw_tarot	One or three cards from the 78-card deck, without replacement
calculate_ziwei	Twelve palaces from a known local birth time and traditional sex parameter
search_library	Search Wenbu guides, symbol notes and reference metadata
read_library	Read a guide or symbol entry by its returned ID
get_daily_card	Public daily reflection and practice; server calendar, no private history

MCP does not call DeepSeek or need your model key. Your host supplies its own interpretation. External reference search results are links and metadata; MCP has no read_reference tool, so the host must open those pages itself before citing their contents. Set locale explicitly to zh or en: MCP defaults to English, while REST and CLI calculations default to Chinese. Some traditional names remain in Chinese.

中文接入说明与可复制示例见 Agent 接入；英文说明见 Agent integrations。

The capability manifest at api/v1/capabilities describes tool conventions and privacy. Tools advertise validated successful-output schemas. Resources include wenbu://methodology, wenbu://knowledge and wenbu://capabilities; selectable prompts are reflect_on_reading and research_with_sources. Prompt retrieval does not call a model.

Try the CLI

The checked-in CLI has no dependencies. This reproducible example uses no personal information and makes no model call:

node public/wenbu.mjs --help
node public/wenbu.mjs iching '{"lines":[7,7,7,7,7,7],"locale":"en"}'

It returns a kind: "iching" JSON result with the six supplied lines and no changing lines. For a new random draw:

node public/wenbu.mjs tarot '{"count":3,"reversals":true,"locale":"en"}'

Keep personal inputs in a file or stdin rather than shell history: node public/wenbu.mjs bazi --file birth.json. The integration page supplies a complete synthetic file. WENBU_URL=http://127.0.0.1:8787 selects a local preview; WENBU_ANALYTICS=off opts out of coarse service analytics.

Read the Skill before adding it to your host. It does not grant access to other conversations or files. A chart's context export and an Agent API request have different schemas; do not pass an export directly to the agent command.

The built-in Agent

Open the workspace or use node public/wenbu.mjs agent --file request.json. This is a separate route that calls DeepSeek on Wenbu's account and requires the person's choice to share the supplied message and context. See the bilingual Agent protocol for a complete request and streaming examples.

The Agent can plan, calculate, read the curated library and available allowlisted pages, ask for missing details and write a report. It has no unrestricted web search or background execution. Reports and their citations are model-authored; a successful source read does not verify every interpretation. Follow-ups preserve the original cards or lines unless a new draw is requested.

Guest conversations and artifacts stay in the browser. Email-verified accounts can enable private cloud history; older browser records require explicit selection for import. Requests send the selected context and a bounded recent history through Cloudflare to DeepSeek. Clearing browser storage removes local records; it does not delete provider records or reset quotas. The app does not promise end-to-end encryption; cloud sync requires an account with cloud history enabled.

Limits and architecture

AI allowance	Current configuration
Single reading	5 attempts per network per Shanghai day
Agent conversation	12 turns per network per Shanghai day
Work per Agent turn	Up to 5 model calls, 12 tool executions and 120 seconds
Site-wide model budget	1,000 attempts per day shared by both AI routes, with Agent use capped at 600

Shared networks can share an allowance. Failed or cancelled attempts may count. Calculators and MCP do not use the model budget, but request rate limits still apply. Visitors are not charged; the operator pays upstream and infrastructure costs.

Cloudflare Workers serves Astro static assets and API routes. A SQLite Durable Object manages AI quotas, D1 holds first-party product events and private feedback, private R2 preserves raw event archives, and rate-limit bindings protect the endpoints. The UI uses Astro 7 and React 19; calculations use pinned versions of lunar-typescript and iztro. Fonts are self-hosted. The optional model runs through DeepSeek's official API with requested model deepseek-v4-flash; responses record the model name reported by the provider.

Feedback is available from every public page and from reading/Agent results. Administrators can inspect linked usage history, triage feedback and download private NDJSON archives at /insights/. See storage, export and privacy details. Personal excerpts are shared only by an explicit choice in the feedback form.

Check and contribute

Public page changes are automatically sent to IndexNow after deployment, with a Cloudflare check every 15 minutes as a fallback. Submission receipts and retries are stored in D1; a receipt is not proof of indexing. See IndexNow operations.

See review evidence, calculation methodology and analytics definitions. Type checks, automated tests, browser checks, deployed smoke tests and actual search traffic are separate kinds of evidence. Software tests do not validate divinatory predictions.

Report a problem with reproducible steps, the selected convention and a synthetic example. Keep birth details, private questions and credentials out of public issues.

MIT license. Third-party packages retain their own licenses.

Email accounts and cloud history

Complete guest trials remain free. Email verification lets users save the current result, sync conversations and journal entries, select older browser records to import, and export or delete their account data. Public MCP and CLI remain available without access to private history. See account architecture, privacy and recovery for limits, reliable saving, quota linking and measurement definitions.