Shotoku

Shotoku checks AI-agent spending against local budgets and policy rules before an action reaches a payment rail or tool. Developers can require approval for selected requests and inspect the resulting decisions in a local audit log.

Shotoku screenshot

Who it's for

  • Developers building TypeScript AI agents that need per-resource spending rules before purchases or paid API calls.
  • Teams using MCP-compatible agents that want policy checks in front of the MCP servers their agents already use.
  • Agent operators who need human review of larger or otherwise selected spending requests through Telegram, a local dashboard, or a terminal UI.

Not the right fit if…

  • Teams looking for a hosted cloud service: Shotoku is described as running entirely on the user's machine, with no cloud service.
  • Teams looking for a custodial payment processor: Shotoku does not hold funds, sign payments, or settle them.

How it fits your workflow

Install the CLI and define policy rules in YAML, including resources, actions, transaction limits, daily limits, and the verdict for a matching request.

Call the TypeScript `authorize` function from an agent, or use MCP support to put checks in front of tool calls. Shotoku returns a structured decision, including reasons and a decision ID.

Decisions are recorded locally in data/decisions.jsonl. Pending requests can be reviewed through Telegram, the local dashboard, or the terminal UI; the agent's own wallet or payment rail handles execution.

Pricing

The vendor doesn't publish prices. Check with Shotoku directly.

Prices checked on Nov 15, 2026 from the vendor's site. They can change; confirm before you buy.

Key features

Budget and policy checks
Configure rules in YAML with per-resource action permissions, maximum transaction amounts, daily limits, and verdicts. A budget can also cap an agent's total daily spending.
Approval and denial decisions
Requests can be approved, denied before payment, or held as pending approval for a human to decide.
Local audit trail
Authorization decisions are recorded in an append-only JSONL file at data/decisions.jsonl, with statuses and decision details.
TypeScript authorization function
Developers can call the `authorize` function from `@shotoku/core` to check an agent action and receive a structured decision with a decision ID and reasons.
MCP integration and wrapping
MCP-compatible agents can use the MCP server, and wrapping mode puts other MCP servers in front of the agent so calls are decided before reaching them.
Human approval interfaces
Pending decisions can be reviewed through Telegram buttons, a local dashboard, or a terminal UI.

Works with

  • Model Context Protocol (MCP)
  • Telegram
  • PayPal

Limitations to know

  • Shotoku is in beta, and the page says its API may evolve.
  • Shotoku does not hold funds, store credentials or private keys, sign payments, or settle them; payment execution remains with the agent's wallet or payment rail.

Alternatives to consider

  • Authoryze

    Choose it if you want AI agents to make purchases using single-use virtual cards while enforcing spending limits and merchant rules.

  • AgentGuard

    Choose it if you also need credential isolation and token metering alongside per-workflow budgets and MCP policy enforcement.

Compare Shotoku side by side

Getting started

npm install -g shotoku-cli
  1. Install the Shotoku CLI using the provided npm command.
  2. Create policy rules in YAML for the resources, actions, limits, and approval verdicts you need.
  3. Integrate `authorize` from `@shotoku/core` into the agent, or connect an MCP-compatible agent.
  4. Review local decision records and handle pending approvals through an available approval interface.

Frequently asked questions

Does Shotoku touch my money or keys?

No. The page says Shotoku holds no funds, stores no credentials or private keys, and never signs or settles a payment. It decides whether a spend is allowed and records that decision; the agent's wallet or payment rail executes it.

Do I need an account or internet connection?

No. Shotoku runs on the user's machine, requires no account, and the page says it has no cloud service or telemetry.

Where are decisions stored?

Decisions are stored in a local append-only file at data/decisions.jsonl inside the project. The page says the file is plain text and can be inspected, moved, or deleted.

Can I use Shotoku with any AI framework or agent?

The page says the core is a TypeScript function usable by an agent that can call a function. MCP-compatible agents can use the MCP server, and wrapping mode can put their existing MCP servers behind Shotoku's policy checks.

Can I approve from my phone?

Yes. `shotoku telegram` sends pending decisions to Telegram with Approve and Deny buttons. The page also describes a local dashboard and terminal UI.

What happens after Shotoku denies a request?

The agent receives a structured response with the denial reason and a decision ID. The page says what happens next is up to the user, such as retrying, surfacing the denial, or halting.

Is Shotoku production-ready?

The page describes Shotoku as being in beta. It says the core authorization path is stable and tested and runs end to end against PayPal's sandbox, but that the API may evolve.