The API

openbooks-api is a Rust HTTP API for double-entry bookkeeping. It’s built for a family or club that wants a real ledger — money in, money out, moved between accounts — without anyone involved needing to know what “double-entry” means.

The stack is axum on top of Postgres, accessed through sqlx. The router and handlers live in openbooks-api/src/lib.rs; src/main.rs is a thin binary that reads config, runs migrations, and calls into it. An MCP server is mounted alongside the HTTP routes so an AI assistant can read the books and record transactions the same way the web UI does; see MCP server.

Running it

Use just from the repo root rather than driving Docker or cargo by hand:

just run      # starts postgres, api, and web in the background
just seed     # loads deterministic sample data (2025 through mid-2026)
just smoke    # end-to-end ledger and report check (needs jq)
just stop     # stop everything

just run waits for Postgres to be healthy before running migrations. Migrations run automatically on API startup (sqlx::migrate!().run(&pool) in openbooks-api/src/main.rs) — there’s no separate migration step to remember.

If you want to run the API directly instead of through just:

docker compose up -d   # postgres on host port 38083
cargo run               # runs migrations, then listens on :38081
./smoke.sh              # needs jq

Logs go to .dev/api.log when run through just; nothing prints to your terminal since the servers run in the background. just status tells you what’s actually up.

Base URL and configuration

The API listens on 0.0.0.0:38081 by default. These environment variables control it:

Variable	Default	Purpose	Read in
DATABASE_URL	postgres://openbooks:openbooks@localhost:38083/openbooks	Postgres connection string	src/main.rs
BIND_ADDR	0.0.0.0:38081	address the HTTP server binds to	src/main.rs
WEB_ORIGIN	http://localhost:38080	the only origin CORS allows to send credentialed requests	src/lib.rs
API_ORIGIN	http://localhost:38081	this API’s own canonical URL, used as the token audience and in the WWW-Authenticate challenge	src/auth/mod.rs
SESSION_TTL_DAYS	30	how long a fully authenticated session lasts from creation	src/auth/session.rs

Every example in this section assumes the default, so requests target http://localhost:38081.

Authentication is required; there’s no RBAC

Every route except GET /health needs a session cookie or a bearer token — see Authentication for the credential types, the /auth/* routes, login backoff, and what phase 1 doesn’t do yet (no second factor, no OAuth). There’s no per-user permission model on top of that: any authenticated user can read and write the whole ledger. That’s deliberate, not an oversight — see the project’s root CLAUDE.md.

CORS is an explicit allowlist

The CORS layer allows exactly one origin — WEB_ORIGIN — with credentials, and only the methods the API actually uses:

CorsLayer::new()
    .allow_origin([origin])
    .allow_credentials(true)
    .allow_methods([Method::GET, Method::POST, Method::PUT, Method::DELETE, Method::OPTIONS])
    .allow_headers([header::CONTENT_TYPE, header::AUTHORIZATION])

The origin is passed as a one-element array rather than a bare value: a bare HeaderValue becomes AllowOrigin::exact, which echoes the configured origin on every response without checking the request’s Origin header at all — not actually an allowlist. The array form is AllowOrigin::list, which does filter, so a request from an unrecognized Origin gets no Access-Control-Allow-Origin header back. CORS governs whether a browser lets script read a cross-origin response, not whether the request gets sent — SameSite=Lax on the session cookie is what actually protects state-changing routes from another site.

Admin commands

User accounts are created and managed on the API binary directly, never over HTTP — anyone who can run these already has DATABASE_URL and could reach the database directly, so there’s no privileged endpoint to protect with a role this system deliberately doesn’t have. Verbatim from admin::USAGE:

openbooks-api                                     serve
openbooks-api user add <email> [--password -]     create a user, print a password
openbooks-api user list                           email, passkey count, last seen
openbooks-api user passwd <email> [--password -]  reset a password
openbooks-api user rm <email>                     delete, cascading sessions/tokens
openbooks-api mint-token <email> [--label <name>] a token for scripts and clients

user add and user passwd print a generated password once — it’s not stored anywhere and can’t be shown again, so write it down or pipe a password in with --password - (reads one line from stdin). user passwd and user rm (via the cascading foreign keys in 0004_auth.sql) end every existing session and revoke every token for that user, the same as a self-service password change through POST /auth/password. user rm also nulls transactions.created_by for that user’s rows — attribution doesn’t cascade, so the transactions stay in the ledger and only the name goes.

From the repo root, just wraps these so you don’t need DATABASE_URL set by hand:

just user-add [email protected]   # user add
just users                      # user list
just mint-token [email protected] # mint-token

See Running the stack for the full recipe list.

Error shape

Handler errors come back as JSON with a single error field:

{ "error": "needs 2+ entries summing to zero (got 1 entries summing to 4500)" }

This is produced by AppError’s IntoResponse impl in openbooks-api/src/lib.rs. Two things map to 400 Bad Request:

• The API’s own validation (missing name, empty description, unbalanced entries).
• A Postgres error whose code is P0001 (a plpgsql raise exception, which is how the balance-check trigger reports a violation) or starts with 23 (a constraint violation, such as an unknown account_id).

Everything else — a real database or server fault — comes back as 500 Internal Server Error with error set to a fixed sentence, “something went wrong on the server”, and nothing more specific: no table or column names, no driver text. The detail is logged server-side (.dev/api.log), not returned — the web app renders error verbatim, and a volunteer treasurer should never see a database’s internals.

Note

The 400 cases above are the ledger’s own invariants rejecting bad input, not the API guessing. The authoritative check is a Postgres trigger described in Data model; the API-level check in create_transaction_data exists only to return a friendlier message before the trigger would fire. A 400 message is about the caller’s own input, and keeps saying so; a 500 is a fault, and stays generic.