Xero MCP

Xero MCP

Covers usage of the Xero MCP server, including authentication (OAuth2 bearer tokens) and related operational notes.

For browser-driven automation of the Xero web UI (not via MCP), see the separate xero-browser skill instead.

MCP vs browser — which to use (READ FIRST)

Before any Xero data task, read references/mcp-vs-browser.md. It is the single source of truth for choosing MCP tools vs the browser Account Transactions report (shared with the xero-browser skill — do not duplicate its content here). Key rule: Xero MCP has no per-account transaction endpoint, so a single account's full ledger comes from the browser report, never from paginating list-manual-journals.

Authentication

The Xero MCP server needs a valid bearer token (XERO_CLIENT_BEARER_TOKEN) in its environment. Tokens last ~30 minutes and must be refreshed via the xero-oauth helper script.

On a 401: STOP, refresh, ask for restart — do NOT limp to workarounds

When any Xero MCP tool returns a 401 / auth error, the correct response is a fixed sequence — not retrying and not falling back:

1. Stop. Do NOT call the tool 2–3 more times hoping it recovers — a 401 is an expired token, it will not self-heal. Do NOT switch to a browser / block-explorer / Cryptio workaround "to keep moving".

2. Refresh the token: run xero-oauth --refresh (writes the new token to .mcp.json / opencode.json / .env).

   NEVER pipe xero-oauth through tail, head, grep, or anything else that crops its output. Run it bare, read ALL of the output. The success/failure indication, and any error, warning, or instruction, are printed across the whole output — crop it and you cannot honestly tell the user whether the refresh worked. Step 3 has you tell the user to restart, which is only sound if you KNOW the refresh succeeded.

3. After successfully running xero-oauth --refresh, your very next output must be one short line asking the user to restart. Then end the turn.

   NO further tool calls. NO saving state. NO notes or beads. NO summary. NO explanation. There are no exceptions — not even work needing no Xero access. The restart kills the process you are running in, so there is no "meanwhile": every extra token delays the restart the user is waiting to perform, and anything you start is discarded.

   "Let me just persist state first" is the most common excuse for disobeying this, and it is wrong. A fresh session re-reads the same files and issue tracker you already have. If state genuinely mattered, it was already saved before the 401.

Why this is the rule, not a suggestion: the MCP path is far faster and more reliable than every workaround (browser Account Transactions report, block-explorer queries, Cryptio cross-checks). Burning turns on retries and then a slow browser workaround is worse on every axis than a 30-second refresh + restart. The ONLY time a browser fallback is right is when the data genuinely isn't available via MCP at all (e.g. a per-account transaction ledger — see references/mcp-vs-browser.md), never merely because the token expired.

A background job caches its token at its own startup; a main-session restart does not refresh a background job's token — kill + re-dispatch it.

Trigger to read references/authentication.md:

• Xero MCP tools return authentication/authorization errors
• Starting a Xero workflow and unsure if the token is current
• Token has expired (~30 minute lifetime)
• First-time setup, or after revoking and re-authorizing the Xero app

That reference covers the xero-oauth script, prerequisites (.env credentials, Xero app registration), the refresh-vs-full-flow choice, and which config files get updated.

Searching for records

The MCP list-* tools paginate at 10 records per page. Absence on page 1 is not absence in Xero. Before concluding a record is missing, search by ID directly or page exhaustively.

See references/searching-records.md for the rules around list-manual-journals, list-invoices, list-bank-transactions, and list-contacts.

Manual journals (create / update / void)

The schema descriptions for create-manual-journal and update-manual-journal are misleading on at least two points:

• lineAmountTypes: "NO_TAX" is rejected by the Xero API (use the default — omit the field — which serialises as NoTax).
• update-manual-journal says "Only works on draft manual journals", but in practice mutates POSTED MJs (narration, date, lines, status).

This makes the void-and-replace pattern (create new POSTED MJ referencing the old UUID, then update the old to status: VOIDED) viable via MCP without touching the web UI.

The response strings from the write tools also render input params verbatim and show line items as [object Object] — so always confirm state with list-manual-journals manualJournalId=<uuid> after every write before trusting it.

See references/manual-journals.md for the full void-and-replace recipe, common error messages, and the token-expiry coupling that bites multi-step MJ workflows.