Reports in plain words
Ask for last month by market or method and get a table you can download as CSV, with the query behind it.
Platform · AI assistant & MCP
Ask your numbers. Approve the next move.
Spark lives in your dashboard and reads the same payments, payouts and orders you do. Ask in plain words, see the data behind every answer, and approve anything it suggests — it never moves money on its own. Prefer your own AI? Connect it through MCP.
Spark · AssistantNorthlight · Wed, Oct 21
Spark: Good morning. I can read Northlight's payments, payouts and orders, and draft changes for you to approve. Ask me anything, or start with a question below.
Proactive insights
It tells you before you ask.
Spark watches approvals, declines, payouts and sales against your own history. When something moves, it writes one line on what changed and why, and links to the screen where you can act.
- Approval drops by market, method and issuer
- Authorizations about to lapse and renewals due
- Payout previews before the money leaves
- Every insight links to the screen it concerns
Spark · InsightsNorthlight · last 7 days
Simulated — insights are written by Spark.
Actions with approval
It drafts. You approve.
Spark can prepare a refund, a price change or a payment link, but it cannot send one. Every draft waits in a queue as a before-and-after, and nothing changes until someone with the right role presses Approve.
- Refunds, payouts, prices and settings always need a person
- Drafts show exactly what changes, field by field
- Approvals follow your team roles
- Drafts and approvals are written to the audit log
Approvals2 waiting
Waiting for you · 2
- 9 min ago
Refund · Owen Reyes
Drafted by SparkFelt Desk Mat · #20297 · pay_1QyV5n
- Status
- SettledRefunded
- Refund
- New:$46.00 to Visa •••• 6610
- Available
- $1,101.47$1,055.47
From Owen's email: "The mat arrived with a torn edge."
- 14 min ago
Price · Ceramic Mug
Drafted by SparkSage glaze · 350 ml · store, links and apps
- Price
- $28.00$32.00
- Starts
- NowMon, Oct 26 · 00:00
- Applies to
- New:Store, links, iOS, Android
You asked: "Raise the mug to $32 from Monday."
Simulated — nothing is sent.
MCP server
Your account, as tools any AI can call.
The SparkIO MCP server exposes your balance, payments, metrics and payment links as Model Context Protocol tools. Add one block to your client's config, paste a restricted key, and your AI can answer from live account data.
- One package: @sparkio/mcp
- Five tools to start, read-only unless you widen the key
- Payment links and refunds are drafts that ask first
- Runs locally over stdio, or remote over HTTP
MCP server@sparkio/mcp
claude_desktop_config.jsonLive key · read + drafts
Connected · 5 tools
- get_balanceAvailable, pending and payoutsRead
- list_paymentsFilter by status, market, method or dateRead
- get_metricsApproval, volume and declines, any windowRead
- create_payment_linkDraft scope: a link you confirmAsks first
- refund_paymentDraft scope: a refund you confirmAsks first
Works with your AI
Claude, ChatGPT, Gemini or Cursor.
Use Spark in the dashboard, or bring your account to the assistant your team already uses. Same tools, same permissions, same audit log — whichever model asks.
- Claude Desktop and Claude Code
- ChatGPT connectors
- Gemini CLI
- Cursor and any other MCP client
Works with your AISame tools · same log
In Claude
Simulated. Brand marks belong to their owners.
In Claude: Summarize refunds and disputes from the last 7 days. It calls list_payments (status=refunded,disputed window=7d). Two refunds: $34.00 to Theo Grant on Mon, Oct 19, for a Paper Tray ordered twice by mistake, and $28.00 to Grace Kim on Fri, Oct 16, for a Ceramic Mug that arrived chipped. No open disputes; the last one, on pay_1QyR2d, was won on Oct 7. Audit log: payments.list · Claude via MCP.
Permissions & privacy
Scoped keys. Read-only by default.
Spark and every MCP client see only what their key allows. New keys can read and nothing else; you widen them one scope at a time, test them on test-mode data, and revoke them in one click.
- Restricted keys, scoped per resource
- Read-only by default; writes become drafts
- Test-mode keys only ever see test data
- Your data is not used to train models
API keysRestricted key
Cursor · MCP
Test moderk_test_7Nq2Vd… · created today by you
Access
- Balances & payouts
- Read
- Payments
- Read
- Metrics
- Read
- Payment links
- None
- Refunds
- None
- Customer contact details
- None
- Send payouts · settings
- Never
Audit log · this key
- 09:40payments.list · status=declinedCursor
- 09:22metrics.read · window=7dCursor
- 08:58key.created · rk_test_7Nq…you
Your data is not used to train models.
Also in Spark
Small things that save an afternoon.
The questions you used to answer with a spreadsheet export.
Anomaly alerts
A message when approvals, declines or refunds move outside their usual range, by email or in the dashboard, with the payments behind it.
Payout forecasts
What lands in your bank and when, from your balance, your payout schedule and the payments still clearing.
Multilingual
Ask in English, Portuguese, Spanish and more; answers come back in your language with amounts in your locale.
From your phone
The dashboard works in any mobile browser, Spark included, so the payout question gets answered from the shop floor.
API access
Call Spark from your own code with the same restricted keys, and get answers back as JSON.
Questions
Questions about Spark and MCP.
- Can Spark move money or change settings on its own?
- No. Spark reads your account and prepares drafts. Refunds, payouts, price changes and settings wait for a person with the right role to approve them, and every draft and approval is written to the audit log with who did it and when.
- Is my data used to train models?
- No. Your payments, customers and conversations with Spark are not used to train models, ours or our providers'. Conversations are kept in your account so you can review them, and you can delete them at any time.
- What can an MCP client see?
- Only what its key allows. A new restricted key reads balances, payments and metrics and nothing else. Customer contact details, refunds and payment links are separate scopes you switch on one by one, and a write scope only ever creates drafts that someone approves.
- Which AI clients work with the MCP server?
- Any client that speaks the Model Context Protocol — Claude Desktop and Claude Code, ChatGPT connectors, Gemini CLI and Cursor among them. Run the server locally over stdio, or point a remote client at the hosted endpoint.
- Can I try it without touching live payments?
- Yes. Test-mode keys only see test-mode data, so you can wire up an agent, ask it anything and approve its drafts without a real payment, refund or payout ever being created.
- What does it cost?
- Spark and the MCP server are included with every SparkIO account. API rate limits apply to MCP calls the same way they apply to the REST API.
Ask your first question today.
Spark is in every dashboard from day one, and an MCP key is one click away.