How to connect Claude to Xero (MCP), step by step

This guide takes about fifteen minutes. Most of that time is in Xero's developer portal. Claude can store the connection and run every query for you; you mint the keys yourself, because keys are never minted over an OAuth connection.

What you need

actionlake is an MCP gateway: one remote MCP server in front of every system you connect. Xero is one of them. The Xero connector covers invoices, contacts, bills, payments, accounts and more, and each action is marked as a read, a write or a destructive change.

Step 1: Add actionlake to Claude

Claude Code. Run one command:

claude mcp add --transport http actionlake https://connect.actionlake.com/mcp

Start claude. On the first tool call that needs a tenant, Claude Code opens the actionlake consent page in your browser. Paste an API key, or press Create one to make a free sandbox tenant on the spot (no signup, no card). Approve, and Claude has eight tools: help, find_actions, get_action, do, ask, check, cost and account.

The Claude app. Open Settings, then Connectors, then Add custom connector. Name it actionlake, paste https://connect.actionlake.com/mcp, add it, then press Connect. The same consent page opens; paste an existing actionlake key there. Creating a new sandbox from the Claude app is not yet tested by us, so if Create one fails there, create the sandbox in Claude Code or the console and paste its key. On Team and Enterprise plans, your workspace owner may need to add the connector.

We tested Claude Code end to end, and the Claude app with an existing key. The connector uses OAuth 2.1 with dynamic client registration, so there is no client id to copy.

Step 2: Create a Xero app and get its tokens

Xero gives API access to an app that a Xero user has approved. Make one app for your business:

1. Sign in at developer.xero.com and create a new app of type Web app.
2. Set a redirect URI that your OAuth tool can receive.
3. Request offline_access (this returns a refresh token) and the accounting scopes your agent needs, such as read access to transactions and contacts.
4. Run the authorisation once. A Xero admin approves the app for the organisation. You receive an access token and a refresh token.
5. Keep the app's client id and client secret. actionlake uses them to renew the access token when it expires.

Creating the app and approving it are the steps a person must do. That is how OAuth works, and it happens once per organisation, not once per agent.

Step 3: Store the Xero connection

Ask Claude: "Connect Xero for me." Claude first asks actionlake what Xero needs. The call is free:

{ "op": "connections.requirements", "service": "xero" }

At the time of writing, the answer asks for four secrets and one setting. Claude then sends the connection:

{
  "op": "connections.create",
  "service": "xero",
  "name": "default",
  "secrets": {
    "access_token": "…",
    "refresh_token": "…",
    "client_id": "…",
    "client_secret": "…"
  },
  "setup_data": { "tenant_scope_id": "your-xero-tenant-id" }
}

The secrets are sealed with AES-256-GCM under your tenant's own data key the moment they arrive. No endpoint returns them, and they never enter a prompt. tenant_scope_id is the Xero organisation to act on. Connected to several organisations? Give each connection its own name; Claude names the one to use on each call.

If you would rather not paste tokens into a chat, enter them in the console instead: actionlake.com/console, then Connections.

Step 4: Give Claude a read-only key

If you pressed Create one in step 1, the key behind Claude is your tenant's root key, and it can do everything. Do not leave an agent on that key for daily work. Mint a narrower one. Keys are minted with an API key itself or in the console (Keys, then New key), never through an OAuth token, so this step is yours:

curl -X POST https://api.actionlake.com/v1/account/keys.create \
  -H "Authorization: Bearer $ACTIONLAKE_ROOT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "claude-xero-reader",
    "allowed_actions": ["xero.read:*"],
    "cap_calls": 2000,
    "confirm_policy": "human"
  }'

• xero.read:* lets the key run every Xero read and nothing else. A write is refused before it reaches Xero. See AI agent permissions.
• cap_calls is a hard limit in upstream calls. At the cap, the next call is refused before it is sent. See agent spending cap.
• confirm_policy: human means a destructive action can never be confirmed by the agent itself.

Disconnect and reconnect the connector in Claude, and paste this key on the consent page. Claude now works under it.

Step 5: Ask Claude about your Xero data

Ask in plain words:

• "Which invoices are overdue by more than 30 days, and how much is owed in total?"
• "List the bills due this week."
• "Find the contact for Acme and show their last three invoices."

Claude finds the right action with find_actions, reads its schema with get_action, and runs it with do. For the first question, the call looks like this:

{
  "action": "xero.invoices.list",
  "input": { "status": "AUTHORISED", "where": "Type==\"ACCREC\"", "order": "DueDate ASC" },
  "limit": 100,
  "intent": "The user asked which customer invoices are more than 30 days overdue."
}

intent is required on every run. It is the agent's own one-sentence reason, and it is stored with the record of the call. See agent intent.

Step 6 (optional): Let Claude write, behind an approval

Reading is where most teams start. When you want Claude to draft invoices or update contacts, widen the key and gate the writes. Send keys.update with your root key, as in step 4, or edit the key in the console:

{
  "key_id": "…",
  "allowed_actions": ["xero.read:*", "xero.invoices.create", "xero.contacts.update"],
  "approval_rules": [
    { "match": "xero.write:*", "approver": "[email protected]" }
  ]
}

Now a do on xero.invoices.create does not run at once. It waits. The finance lead gets an email with a signed link, or decides in the console, and the call runs only after they approve. Under confirm_policy: human, a write that creates or changes a billing record, such as an invoice, also needs an explicit confirm: true. A destructive action, such as xero.invoices.void, needs a person's approval: the agent cannot confirm it on its own. See human-in-the-loop approval.

Before anything runs, Claude can call check with the same payload. It validates the input, the key's permissions and the connection, and says what the call will cost. It is free and sends nothing to Xero. See dry run.

What it costs

actionlake bills the upstream call: one HTTP request sent to Xero that returned a definitive answer.

The Free plan includes 100,000 upstream calls a month and stops there, so it cannot overrun. Paid plans start with Starter at US$49 a month for 250,000 calls. See pricing.

Several Xero organisations

Accountants and bookkeepers often look after many organisations. Connect each Xero organisation as its own named connection, each with its own tenant_scope_id. The agent names the connection, and with it the organisation, on every call, and the record shows which organisation each call reached.

Xero's own MCP server, or a gateway?

Xero publishes its own MCP server, @xeroapi/xero-mcp-server (XeroAPI on GitHub, checked 22 October 2026). Your MCP client launches it on your own computer, with a Xero app's client id and secret, or a bearer token, in its settings. For one person using Xero in one client, that may be all you need.

Xero plus your other systems

The reason to reach Xero through a gateway, rather than a Xero-only server, is everything else: one hosted URL for every agent, a key per agent with a cap and approval rules, a record of every call, and your other systems behind the same URL. The same tenant can hold your PSA, CRM or payroll. Claude can then ask one question across two systems with ask, and only the answer comes back. Guides: joining Autotask tickets with Xero invoices and MCP server vs MCP gateway.