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.

Related MCP server: enterprise-agent-lab

What's inside

src/registry.js            declare once; derive tools, forms, guards, audit; enforce
src/workspace.js           a SQLite database per space (per person, per sandbox)
src/members.js             capability-token identity: named members with roles
src/oauth.js               OAuth 2.0 front door for MCP clients (DCR, client-id metadata documents)
src/mcp-http.js            MCP over streamable HTTP; one tool list per space's modules
src/document.js            the invoice document: one PDF renderer, previewed as its own pages
src/modules/core, o2c, crm, prospect, jobhunt, solo, purchases, bridge
specs/<area>/spec.md       the area spec: calibration, entities, invariants, read models
specs/<area>/scenarios/    executable conformance scenarios (refusals are contract)
src/conformance.js         replays scenarios through the real registry; keeps evidence
server.js                  workbench + hosted demo + MCP + OAuth (all one registry)
mcp-server.js              stdio MCP for local development
test/contract.test.js      the 24-gate contract (see below)
test/smoke.mjs             every workbench view and detail panel rendered in headless Chrome
test/local/                plain-language scripts run against a local model, checked in the database

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

npm install
npm test          # the 24 gates
npm run smoke     # every workbench view and panel, in headless Chrome
npm run local -- --model <id> --script plain-invoicing --instructions off   # a local model, plain prompts
npm run demo      # a full quote-to-cash run, human and agent interleaved
npm start         # workbench on http://127.0.0.1:8140
npm run mcp       # stdio MCP server (OTC_WORKSPACE=you)

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:

{ "mcpServers": { "saybooks": {
    "command": "/usr/local/bin/node",
    "args": ["/path/to/saybooks/mcp-server.js"],
    "env": { "SAYBOOKS_WORKSPACE": "main", "SAYBOOKS_MODULES": "core,solo", "SAYBOOKS_ACTOR": "gemma-4" }
} } }

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.