Skip to main content
Okaro for Agents
Beta

Use Okaro inside Claude & ChatGPT

Add Okaro as a connector in Claude or ChatGPT to log an expense in one sentence, scan receipts, or ask “how much did I spend this month?”. Or plug the local MCP server and CLI into Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Codex and other agents. Sign in with your own Okaro account; the AI only sees your data.

Remote MCP
Claude · ChatGPT · Connectors
Local MCP
Claude Desktop · Cursor · Claude Code
CLI
npm · Node.js 22+ · --json
Sign-in
Google · Apple · Email

Use Okaro inside Claude & ChatGPT

Beta

Nothing to install: add the Okaro connector (a remote MCP server) in Claude or ChatGPT, sign in to your Okaro account and allow access — then log expenses right in the chat. Currently in Beta.

MCP server URL

https://okaro.harukibox.com/mcp

Claude

claude.ai on the web, desktop and mobile apps. Custom connectors need a Pro, Max, Team or Enterprise plan.

  1. Open Settings → Connectors and click “Add custom connector”.
  2. Name: Okaro. URL: https://okaro.harukibox.com/mcp. Click Add.
  3. Click Connect, sign in to your Okaro account, and click Allow on Okaro’s consent page.
  4. Back in a chat, turn on Okaro from the tools menu (Search and tools).

ChatGPT

Needs a plan that supports developer mode (e.g. Plus, Pro, Business).

  1. Open Settings → Apps & Connectors → Advanced settings and turn on Developer mode.
  2. Back in Apps & Connectors, click “Create”.
  3. Name: Okaro. MCP Server URL: https://okaro.harukibox.com/mcp. Authentication: OAuth.
  4. Check “I trust this application”, click Create, sign in to your Okaro account and click Allow on Okaro’s consent page.
  5. In a chat, pick Okaro from the “+” / Developer mode tools menu.

Once connected, try

  • Log it: PX Mart today, NT$368
  • (Upload a receipt photo in ChatGPT) Log this receipt — my card was actually charged NT$2,752
  • How much did I spend this month? How does it compare with last month?

The connector has the same tools as the local MCP server, minus sign_in / sign_out (the connection itself is the sign-in), and it never reads local files: receipt photos you upload in ChatGPT are passed to Okaro automatically, or you can give an image URL. It can only access your own Okaro data (isolated by Supabase RLS), with the same plans and free-plan limits as the app. Disconnect it in Claude or ChatGPT settings at any time to revoke access.

Beta

The Claude / ChatGPT connector, MCP server and CLI are in Beta — use them with your Okaro account

The connector (remote MCP server) needs nothing installed: add its URL in Claude or ChatGPT and sign in to Okaro. The local MCP server (@harukibox/okaro-mcp-server) and CLI (@harukibox/okaro-cli) are open source under MIT and require Node.js 22+. All of them use your own Okaro account — no separate keys to request. They are still in Beta, so tool names and setup steps may change.

Why agents pick Okaro as their bookkeeping backend

Native receipt OCR

Gemini 2.5 returns structured JSON for amount, merchant, date and line items, so your agent skips the OCR plumbing. record_receipt scans and saves in one step.

Original amount and actual charge, kept apart

Every transaction keeps its original amount and currency plus a settlement amount in your settlement currency. Foreign purchases need the amount actually charged or an exact rate you give — Okaro never guesses it from reference rates. convert_currency only returns reference conversions, labelled as estimates.

Your own account, isolated by RLS

Run okaro login with your Okaro account (Google, Apple or email). The agent acts as you, and Supabase RLS means it only sees your data. okaro logout revokes just the CLI / MCP session — the web app and iPhone stay signed in.

Dual-receipt reconciliation built in

Each transaction can keep both the supplier receipt and the payment proof, so agents can help with reimbursement, tax filing and customer billing.

Three surfaces — pick one, or combine them

The Claude / ChatGPT connector for bookkeeping right in the chat, with nothing to install. The local MCP server for desktop agents like Claude Desktop and Cursor. The CLI for shells, scripts and Claude Code. The website itself also supports WebMCP, so in-browser AI agents can use Okaro’s site tools in supporting browsers.

Claude / ChatGPT connector

Beta · sign in with Okaro

A remote MCP server (Streamable HTTP) with OAuth 2.1 sign-in to your Okaro account (dynamic client registration and PKCE). Add the URL as a Claude custom connector or a ChatGPT developer-mode app. Same tools as the local server, minus sign_in / sign_out.

https://okaro.harukibox.com/mcp
See how to connect →

Local MCP server

Beta · sign in with Okaro

Speaks the Model Context Protocol, no environment variables needed. Works with Claude Desktop, Cursor, Claude Code, VS Code, Windsurf, Codex and other MCP clients; if you are not signed in, the sign_in tool opens Google / Apple sign-in.

npx -y @harukibox/okaro-mcp-server
See the quickstart →

Okaro CLI

Beta · sign in with Okaro

Commands like `okaro login`, `okaro tx add`, `okaro receipt add`, `okaro summary`, `okaro fx` and `okaro export`. Every command supports --json, so it fits shell pipelines and scripts.

npm i -g @harukibox/okaro-cli
See the quickstart →

Tools available to your agent

The main MCP server tools. Every tool runs as the signed-in user, follows the same validation and free-plan limits as the web and iOS apps, and returns structured JSON. Through the Claude / ChatGPT connector the tools are the same, except there is no sign_in and receipts come from uploaded images or image URLs.

ToolDescription
get_accountSign-in status, plan, settlement currency and usage.
sign_inSign in to your Okaro account with Google / Apple in the browser (local MCP server only; the connector signs in via OAuth).
record_receiptScan a receipt and save it as a transaction in one step.
analyze_receiptReceipt image → structured JSON (amount, merchant, date, line items), without saving.
create_transactionCreate a transaction: original amount and currency, settlement amount, category and card.
list_transactionsQuery transactions by date, merchant, category, card or currency.
update_transactionEdit an existing transaction’s amount, category, card and more.
list_cardsList the user’s payment cards, optionally with this month’s spend.
summarize_periodTotals, change vs. the previous period, and top categories / cards / merchants, grouped per settlement currency.
convert_currencyReference conversion from public daily rates (labelled as an estimate, not used for bookkeeping).
export_transactionsExport CSV / JSON (Premium).

Local MCP server: wire your agent up in 60 seconds

Claude Desktop / Cursor / Claude Code example (requires Node.js 22+):

Step-by-step beginner guide (Chinese) →
  1. 1. Sign in to your Okaro account

    Choose Google, Apple or email + password; Google / Apple finish in the browser. Credentials are stored in ~/.okaro/credentials.json and shared by the CLI and MCP server. For headless setups, set OKARO_EMAIL and OKARO_PASSWORD instead.

    npx -y @harukibox/okaro-cli login
  2. 2. Add the MCP server

    Add this to your MCP client config (Claude Desktop, Cursor, …) — no environment variables needed. In Claude Code, run claude mcp add okaro -- npx -y @harukibox/okaro-mcp-server. okaro mcp --client cursor prints the snippet for each client.

    {
      "mcpServers": {
        "okaro": { "command": "npx", "args": ["-y", "@harukibox/okaro-mcp-server"] }
      }
    }
  3. 3. Invoke from chat

    Tell Claude / Cursor "log this Don Quijote receipt onto my reseller card — I was charged 2,752". It will call record_receipt, or chain analyze_receipt → create_transaction.

Register your agent on the Okaro Agent Directory

The Okaro Agent Directory is a curated public directory shown to every Okaro user. Submit your agent name, purpose and use case — once approved, your agent gets listed and earns referral revenue share.

Submit your agent

FAQ

How do I use Okaro in ChatGPT?+

In ChatGPT, open Settings → Apps & Connectors → Advanced settings and turn on Developer mode. Back in Apps & Connectors, click Create; set Name to Okaro, MCP Server URL to https://okaro.harukibox.com/mcp and Authentication to OAuth, check I trust this application, click Create, then sign in to Okaro and click Allow. In a chat, pick Okaro from the “+” / Developer mode tools menu. You need a ChatGPT plan that supports developer mode.

How do I use Okaro in Claude?+

On claude.ai (web, desktop or mobile), open Settings → Connectors, click Add custom connector, set Name to Okaro and URL to https://okaro.harukibox.com/mcp, click Add, then Connect, sign in to Okaro and click Allow. In a chat, turn on Okaro from the tools menu (Search and tools). Custom connectors need a Claude Pro, Max, Team or Enterprise plan.

What data can the connector see, and how do I revoke access?+

Only your own Okaro data: the connector acts as the Okaro account you authorised, Supabase RLS keeps everyone else’s data out of reach, and the same free-plan limits as the app apply. Disconnect or remove Okaro in your Claude or ChatGPT connector settings at any time; access tokens also expire on their own, after which you authorise again.

How does a user authorise my agent?+

When a user adds the Okaro connector in Claude or ChatGPT, Okaro’s consent page opens; they sign in and click Allow (OAuth 2.1). For the local tools, they sign in to their own Okaro account with okaro login (or the MCP sign_in tool). No API keys either way. The agent acts as that user and Supabase RLS limits it to their data; transactions written from the CLI / MCP are tagged with their source. okaro logout revokes the CLI / MCP session, while the web app and iPhone stay signed in.

Are the MCP server and CLI open source?+

Yes. @harukibox/okaro-mcp-server and @harukibox/okaro-cli are MIT-licensed and currently in Beta.

Does Okaro support WebMCP?+

Yes. When a browser exposes the W3C WebMCP API (document.modelContext / navigator.modelContext, currently experimental in Chrome), okaro.harukibox.com registers tools for in-browser AI agents. Anyone can use okaro_about, okaro_navigate (opens site pages) and okaro_agent_setup (MCP / CLI setup instructions per client); after signing in to the web app, get_account, list_transactions, summarize_period, list_categories, list_cards and create_transaction are also available. The tools run with the browser’s own login session and data stays RLS-isolated. For foreign-currency purchases, create_transaction requires the settlement amount actually charged, and it is marked as consequential so the browser asks before running it. In browsers without WebMCP, nothing happens.

Can my agent upload receipt images on behalf of the user?+

Yes. With the local MCP server, analyze_receipt and record_receipt accept local file paths, image URLs or base64; through the connector they use photos uploaded in ChatGPT (passed to the tool automatically) or image URLs. Receipts are stored in the user’s own Supabase Storage under RLS isolation.

Is there a rate limit?+

Yes: receipt analysis is limited to 10 requests per minute per user, and export to 10 per minute. Free-plan limits match the web and iOS apps: 50 transactions in total, 3 custom categories, 1 card, and export is Premium-only.

Can my agent earn revenue?+

Yes. Agents listed on the directory receive 20% recurring share when a referred user upgrades to Premium.

    Okaro for Agents — Track expenses inside Claude & ChatGPT | Okaro