Saybooks
io.github.alviso由 GitHub 账号 alviso 发布v0.4.0更新于 Nov 3, 2026
你的AI根据你自己的银行对账单记账:银行卡、贷款、账单、收据和纳税年度。
AI 翻译由 AI 模型根据下方的原文描述翻译,可能有误。
原文描述
Books your AI keeps from your own bank statements: cards, loans, bills, receipts and the tax year.
概览
Saybooks 的托管 MCP 服务(saybooks.io);需要 OAuth 登录;Streamable HTTP。
暂无 AI 概览:以下信息来自注册中心条目和服务本身。
- 端点
- saybooks.io
- 最近检查
安装
在 SourceWeft 中
- 打开 控制台中的 Saybooks,将其添加到工作区。
- 连接 所需的 凭据。它们会被私密存储,绝不会在公开页面上展示。
- 为需要使用其工具的对话启用该服务。
Web executable,通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。
交给智能体安装
把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体,或 SourceWeft 对话中。智能体会读取该服务器的安装指南,先向你展示发布者、端点和登录方式,经你同意后再把它添加到你的客户端。密钥不会写进任何命令。
阅读 https://sourceweft.com/mcp/io.github.alviso%2Fsaybooks/install.md,并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前,先告诉我它的发布者、端点和登录方式,等我确认。不要把 API 密钥写进命令或我会提交的文件;问我密钥应从哪里获取。其他 MCP 客户端
把它添加到你客户端的 mcpServers 配置中。 客户端会在首次连接时完成 OAuth 登录。
{
"mcpServers": {
"saybooks": {
"type": "http",
"url": "https://saybooks.io/mcp"
}
}
}README
Saybooks
The books you can talk to. The books that stay books.
Saybooks is an open-source ERP where your agent and your interface are the same system: one command registry, one rulebook, one audit trail. Tell Claude to import a card statement, issue the invoice or ship the order, or click the button: same command, same rules, same record of who did what and why.
For your household: saybooks.io/personal. Hand your agent
the bank and card statements and it keeps the accounts, the card, the mortgage and the car
loan, the bills and the receipts, with a year summary for taxes. No bank feed and no bank
password: a statement is taken only when its rows add up to its printed balances. Any bank,
any currency. See a made-up household with no signup,
connect https://saybooks.io/mcp/household, or watch the five-part series.
Live: saybooks.io. Try the demo with no signup, or sign in with Google
and pick the modules your books carry. Connect Claude by adding https://saybooks.io/mcp as a
connector (OAuth, no keys to paste); the same books are in the browser as ordinary pages.
Watch: Saybooks process demos, five short videos recorded live against the demo books: purchase orders and receiving, inventory value, the month-end close, bank reconciliation, and sales tax.
Fourteen modules today, every one spec-first with executable conformance scenarios:
Why this exists
"Conversational ERP" usually means a chat layer bolted onto a system that doesn't know the agent exists. Saybooks inverts that: every business act is declared once, and everything else is derived from that declaration:
Nothing outside the registry may call a handler. A click and a tool call go through the same
command, the same guards, the same transaction. The two surfaces cannot drift, because neither
is authored, and test/contract.test.js fails the build if anyone tries.
Agent-first, not parser-first: where a file is involved (a statement, a receipt), the agent reads it and hands over what it read; the module checks the control totals, keeps the provenance of every row, and refuses what does not reconcile. No bank parsers to maintain.
What's inside
Access control: every command declares a permission tag; four roles (owner / controller / clerk / viewer); enforcement at the registry choke point; denials are one-sentence refusals, logged. An agent connecting through a member's key is that member's delegate: their name, their permissions. Every write carries a reason; reads are never logged.
Money is integer minor units everywhere, with a currency; sums are per currency and never cross. Nothing is invented: a missing amount, date or name is asked for, never guessed.
The 24-gate contract
Every module — present and future — is held to: MCP/UI parity · namespace prefixes · doctrine on every write · guards declared · table ownership (no cross-module writes) · a tool budget · module mounts · one-sentence rule · audit behavior · fixture replay · spec conformance (an implementation claiming an area must map every act and pass every scenario) · permissions (unpermissioned commands do not ship) · spec version (a module implements the version its acts file declares) · human-only (a field or a whole act a person must do, refused to agents whatever their role) · the door (staged rows are judged by an agent and promoted only by a person) · the claim gate (what a campaign may say is a person's to set, and holds on edit) · bills (only a person approves paying one; a payment is recorded after it happened, against a statement row that matches it).
The endgame is competing implementations of the same area, certified by replaying the same scenario files — the spec speaks in acts, not commands, so any conforming module runs them.
Run it
Register for a Claude Code session: claude mcp add saybooks-local -e SAYBOOKS_WORKSPACE=$USER -- node <path>/mcp-server.js
Local models
Saybooks is a plain MCP server, so a model on your own machine can keep the books with nothing
leaving it. Tested with Gemma 4 26B A4B in LM Studio on a 32 GB laptop: a full invoice cycle and
a full statement cycle from plain sentences, no coaching. In LM Studio's mcp.json:
SAYBOOKS_MODULES picks the tools the model sees (core,solo is 35 tools; core,purchases 60;
all fourteen modules is 236 and too many for a small model). Load the model with one concurrent slot.
What that took, and the six fixes it forced in the tools, is in
saybooks.io/notes/local-models; the plain-language
harness that measures it is npm run local.
Honest limits
Any currency per invoice but no conversion between them (sums stay per currency), one company
per space, no tax engine (rates are captured and frozen per line; determination is an
integration's job), no outbound anything: Saybooks never emails a customer, charges a card, or
books a shipment. It records what happened. The ledger hand-over writes import files for
QuickBooks Online and Xero; it does not post over their APIs. Each area's deferred list, with
reasons, is in its spec under specs/.
License
AGPL-3.0-only. Copyright (C) 2026 Peter Varga. If you run a modified Saybooks as a service, share your changes — that's the deal.
来源:README.md,提交 b14f4cd
工具
0版本历史
1- v0.4.0最新Nov 3, 2026
评分
已满足适用检查项的 85%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通必需25 分
- 已声明认证方式必需10 分
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分