Open-source spend-control layer that sets budgets, routes big AI agent spends to human approval, and logs every decision
Shotoku is an open-source spend-control layer for AI agents. Before your agent pays for something, it asks Shotoku. Shotoku checks the request against rules you write in one YAML file and approves it, denies it, or holds it for you to review. It runs locally, holds no funds or keys, and logs every decision. It's built for developers whose agents call paid APIs, use MCP tools, or make purchases, and who want limits that the agent can't approve for itself.
policy.yaml file. Rules are checked top to bottom, and the first match decides, so you can always tell why a request was approved, denied, or held.maxAmount), rolling 24-hour per-vendor caps (maxDailyAmount), and an overall daily budget per agent. This stops a runaway loop from draining your account.shotoku ui), or on your phone through Telegram's Approve and Deny buttons.authorize() from @shotoku/core to add spend checks to agents that don't use MCP.npx shotoku-cli demo before installing.shotoku telegram to get pending requests on your phone, then tap Approve or Deny without returning to your desk.authorize() from @shotoku/core before each paid action, and you get policy checks and an audit trail without running MCP.Shotoku is a spend-control layer for AI agents. Your agent asks Shotoku before it spends money. Shotoku checks the request against your YAML policy and approves it, denies it, or holds it for a person to approve.
Yes. The Open Source edition is free. It includes the authorize() core and policy engine, the hash-chained decision ledger, the CLI and terminal UI, the MCP server, x402 payment authorization, and signed snapshots. A Team plan is in development, and Enterprise pricing is available on request.
No. Shotoku holds no funds, stores no credentials or private keys, and doesn't sign or settle payments. Your agent's own wallet or payment rail carries out approved payments.
It works with any MCP-compatible agent, including Claude Desktop, Claude Code, and Cursor. Agents that don't use MCP can call the authorize() function from the @shotoku/core TypeScript package.
In basic advisory mode, yes, because the agent has to choose to ask. To enforce checks, put Shotoku in front of your other MCP servers so every tool call goes through it first.
0 out of 5 stars
Based on 0 reviews
5 star reviews
4 star reviews
3 star reviews
2 star reviews
1 star reviews
If you've used this tool, share your thoughts with other users
Give your AI agents a budget, not your card
Cited physical-world data and tools for AI agents
Give your AI agent its own email inbox
Give your AI agent an @address to find deals
AI voice agents that make real phone calls
Connect AI tools to your work apps safely
Live SEO and business data inside your AI agent
AI-powered financial close for accounting teams
Long-term memory for your AI chats