wanderbooks/docs/mcp-interface.md

MCP interface spec — ledger-mcp

Goal

The primary interface to the books is an AI agent driving MCP tools. The engine (ledger-core) stays pure, ledger-store persists, and ledger-mcp exposes both over the Model Context Protocol — the open standard, so the books are drivable by any MCP client (Muse, Claude Code, IDE agents), not locked to one assistant.

Crates in this phase

Crate	Role	Spec
ledger-store	SQLite persistence, blob metadata, event log	docs/receipt-store.md
ledger-mcp	MCP server wrapping store + core	this document

Rust, rmcp SDK (the Rust MCP SDK). The MCP crate holds no accounting logic: it translates tool calls into Ledger commands and formats results. If a behavior can't be expressed through ledger-core's public API, the core is missing something — fix the core, don't work around it in the MCP layer.

Transports

1. stdio (build first). For Claude Code, Claude Desktop, local agents. No auth — the OS user boundary is the auth.
2. Streamable HTTP (later in the phase). For remote access (phone, travel laptop). Bearer token, binds 127.0.0.1 or the tailnet interface only. Binding 0.0.0.0 is an explicit human decision, never a default.

Server instructions

The MCP initialize response carries an instructions string: the agent operating contract (below) in condensed form. The model reads it at the start of every session, before any tool call. Tool descriptions repeat the relevant fragment at the point of use (post_entry's description states the preview-before-post rule; attach_receipt's states the receipt rule). Prose guidance is therefore delivered twice — once at session start, once at the tool — but it is guidance, not enforcement. Enforcement is the receipt policy below.

Tool design principles

• Narrow tools, explicit arguments. No "do what I mean" tools. The agent supplies dates, accounts, sides, amounts; the server validates.
• Amounts are decimal strings in requests and responses. Never floats, never locale-formatted numbers.
• Every mutating tool echoes the exact resulting state. The "show the numbers" rule from AGENTS.md applies over the wire too.
• Preview-before-post. post_entry and reverse_entry take confirmed: bool. Without confirmation they return a preview (full lines
  • resulting account balances) and post nothing. The agent must present the preview to the human before confirming. This encodes the standing fact-check rule in the protocol itself.
• Errors are typed strings mirroring LedgerError variants, with the offending values included (e.g. unbalanced totals).

Tools — books

• books.chart → { accounts: [{ id, code, name, type, parent, placeholder, balance }] }. Balances included: one call answers "what do we have."
• books.draft_entry { date: "2026-10-10", description: "..." } → { entry_id, status: "draft" }.
• books.add_line { entry_id, account: "1000" | <id>, side: "debit" | "credit", amount: "123.45" } → { lines: [...], totals: { debits, credits, balanced } }. Running totals after every line: the agent watches the entry balance in real time.
• books.post_entry { entry_id, confirmed } → without confirmation: { preview: { lines, resulting_balances, receipts: { attached: [...], required_but_missing: bool } } }; with confirmation: validates (including the receipt policy), posts, returns { entry, balances_changed: [{ code, before, after }] }.
• books.reverse_entry { entry_id, date, description, confirmed } → same preview/confirm pattern as post.
• books.get_entry { entry_id } → full entry with lines.
• books.search_entries { account?, date_from?, date_to?, text?, limit?, offset? } → entries with lines, newest first. limit caps at 100.
• books.trial_balance {} → canonical JSON (same shape as reports::snapshot trial-balance section).
• books.balance_sheet {} → canonical JSON.
• books.profit_and_loss {} → canonical JSON.
• books.digest { since: "2026-10-01" } → { entries_posted: [...], totals_by_account: [...] }. The human review surface: "what did the AI do with my books lately." The digest is the answer to "trust, but verify."

Tools — receipts & documents

• books.attach_receipt { entry_id?, file_name, mime_type, content_base64 } → { receipt_id, sha256, deduplicated, size_bytes }. entry_id is optional: a receipt can be filed before its entry exists.
• books.link_receipt { receipt_id, entry_id } → link a previously filed receipt to its entry.
• books.list_receipts { entry_id? } → metadata list (no bytes).
• books.get_receipt { receipt_id } → { metadata, content_base64 }.
• books.file_document { category: "formation" | "legal" | "tax" | "accounting" | "other", file_name, mime_type, content_base64, description? } → { document_id, sha256, deduplicated }. The Drive-folder replacement (today's 01 Formation / 02 Legal / 03 Tax / 04 Accounting).

Receipt enforcement

Prose tells the model to attach receipts; the policy makes it true. The engine stays policy-free (it knows nothing of receipts) — enforcement lives in the MCP layer, which reads the attachments table before confirming a posting.

• Policy config (server config file, not per-call): require_receipts_above (default "25.00" in base currency), plus receipt_required_accounts (account codes that always require receipts, e.g. meals and travel) and receipt_exempt_accounts (e.g. automated bank fees). Human-edited and versioned with the deployment.
• Preview. Every post_entry / reverse_entry preview includes the receipt checklist (attached: [...], required_but_missing: bool), so the human sees it before confirming.
• Hard error. Confirming a posting that violates the policy fails with a typed error: { "error": "receipt_required", "entry_id": …, "missing": …, "policy": "require_receipts_above = 25.00 USD" }. The agent's job then is to get the receipt attached (ask the human for it) and re-confirm — not to shrink the amount or split the entry to dodge the threshold. The tool description says this explicitly; gaming the policy is a correctness violation, and the append-only audit log would show it anyway.
• Below the threshold, receipts are still expected per the operating contract ("a material entry without its source document is incomplete") but not hard-required — the preview checklist keeps them visible.

Tools — export

• books.export { format: "csv" | "json", scope: "journal" | "trial-balance" | "profit-and-loss" | "balance-sheet" | "chart" } → { content_base64, file_name }. CPA-ready output. xlsx and iif (QuickBooks Desktop journal entries) arrive with R-FUTURE-4 via the ledger-export crate.

Resources (read-only)

• books://chart
• books://reports/trial-balance, books://reports/balance-sheet, books://reports/profit-and-loss
• books://digest?since=2026-10-01
• books://receipts/{id} — receipt bytes with metadata

Resources exist so read-only clients (dashboards, the future read-only GUI, auditors) never need tool credentials.

The agent operating contract

Any AI client driving these tools follows this contract (it will also live in the server's tool descriptions, so the model reads it at use time):

1. Show the preview before confirming any posting. No silent posts, ever.
2. Echo exact numbers back to the human. Amounts as written, no rounding in prose.
3. Attach receipts at record time. A material entry without its source document is an incomplete entry. File first, link later if the entry isn't ready. Postings at or above the receipt policy threshold cannot post without one — the server refuses with receipt_required.
4. When unsure of categorization, draft and ask. A draft is cheap; a wrong posting is a reversal. Never guess-post.
5. The digest is the receipt for your work. After a batch of bookkeeping, offer the digest; the human reviews, you fix.

Out of scope for this phase

Bank feeds, payment processing, payroll, tax filing, multi-user authN/Z, the GUI (it consumes the event log in a later phase). Cash flow reporting is R-FUTURE-3 and rides the engine roadmap, not the MCP layer.