FrankKi MCP

FrankKi MCP

Hosted remote MCP server that lets AI agents compose, price, send, and track real physical letters worldwide.

Category
访问服务器

README

<!-- Repo: frankki-mcp | Homepage: https://frankki.app/mcp | Topics: mcp, mcp-server, claude, ai-agents, physical-mail, letters -->

FrankKi MCP

Hosted, remote MCP server. Nothing to install, nothing to self-host. Point your agent at https://mcp.frankki.app (remote, streamable-http, OAuth 2.1 or API key). This repository is the public manifest, documentation and changelog for that server; the server itself is closed-source and runs on our own infrastructure, which is why there is no src/ directory here.

Physical mail for AI agents. With one tool call, FrankKi MCP lets an agent send a real, printed letter to a physical address anywhere in the world. Compose, price, send, and track. The document is printed, franked, and delivered by postal partners worldwide.

This is the official Model Context Protocol (MCP) server for FrankKi. It is built for B2B and developer use: give your agents, backends, and automated workflows the ability to put real letters in the mail, at scale, worldwide.

Website: frankki.app · MCP for humans and agents: frankki.app/mcp (English: frankki.app/en/mcp)

Status MCP Delivery Transport Auth


Contents


What is FrankKi MCP?

FrankKi MCP is the agent-native interface to FrankKi's physical mail infrastructure. FrankKi composes, prints, and physically mails letters. This server exposes that capability to software: your model drafts a letter and sends it, and a real document arrives at the recipient's address.

Delivery is worldwide, handled by our postal partners in each destination country. It is primarily a B2B and developer product: if you are building an assistant, an ops automation, or a backend that needs to produce and dispatch physical correspondence, this is your physical-output layer.

Why physical mail via MCP?

Some things still have to arrive on paper: formal and legal notices, transactional and compliance mail, contract confirmations, dunning and reminders, customer onboarding, official correspondence. These workflows usually dead-end at "now a human prints this and walks to a postbox." FrankKi MCP removes that step. The same agent that drafts the letter can send it, worldwide, and track delivery.

Typical uses:

  • Dunning and payment reminders
  • Cancellations, terminations, objections, and other formal notices
  • Contract and order confirmations
  • Customer and account notifications
  • Onboarding and welcome letters
  • Any workflow whose last step must be a physical letter

What your agent can do

The server exposes a focused, scoped tool set. The core send flow:

Tool What it does Side effect
address_validate Validate a recipient address against destination-country postal rules read only
address_search_company Look up a company's postal address by name read only
template_list Browse reusable letter templates read only
letter_create_draft Draft a professionally formatted letter from plain text plus a recipient, returns a preview creates a draft, no send
shipping_quote Get the exact price for a send to a given destination before committing read only
wallet_balance Read the prepaid wallet balance read only
order_send Send one physical letter, charges the wallet; supports a maxCostEuros guard spends money, mails paper
order_send_batch Bulk send: one call, many recipients with merge fields spends money, mails paper
approval_submit Route a letter into the four-eyes approval queue instead of sending directly queues, no send
order_status Track a letter through printed, posted, and delivered read only
wallet_topup_link Create a secure Stripe top-up link creates a payment link

Beyond the core flow there are tools for attachments, letterheads and signatures, sender profiles, send presets, scheduled sends, order cancellation, posting receipts (Einlieferungsbeleg), client (Mandanten) lookups, and GoBD/DATEV archive exports. The full list with scopes and side effects: docs/tools.md. The authoritative, versioned list is always what the server returns from the MCP tools/list method.

Composition is text-first: you provide the letter content and the recipient, and FrankKi produces the correctly laid-out document (DIN 5008 by default). Attachments are supported. Sending a pre-rendered arbitrary PDF is intentionally not part of the platform.

Quick start for humans

You do not clone or self-host anything. The FrankKi MCP server is hosted. You point your MCP client at it and authenticate.

1. Get access. Create a partner account and issue a key from the dashboard at frankki.app/mcp. Fund your wallet, or use the sandbox first (no real mail).

2a. Claude Desktop and clients with native remote support. Add a custom connector pointing at the endpoint. OAuth runs in your browser on first use:

{
  "mcpServers": {
    "frankki": {
      "url": "https://mcp.frankki.app"
    }
  }
}

2b. Clients that only speak stdio (older Claude Desktop, some Cursor setups). Bridge with mcp-remote:

{
  "mcpServers": {
    "frankki": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.frankki.app"]
    }
  }
}

2c. Headless with an API key (no interactive browser). Pass the key as a bearer header:

{
  "mcpServers": {
    "frankki": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote", "https://mcp.frankki.app",
        "--header", "Authorization: Bearer ${FRANKKI_API_KEY}"
      ]
    }
  }
}

2d. Claude Code. One command:

claude mcp add --transport http frankki https://mcp.frankki.app

Restart the client, and your model can now send letters. Ask it something like: "Send this payment reminder to the address on file, but show me the price first."

Step-by-step guides for Cursor, VS Code, Windsurf, Zed, the OpenAI Agents SDK, and LangChain: docs/clients.md.

Quick start for AI agents

The server is standard streamable HTTP MCP. Discovery, endpoint, and auth in one place:

  • Endpoint: https://mcp.frankki.app
  • Registry name: app.frankki/letters (published in the Official MCP Registry, version 1.0.0)
  • Auth metadata: https://mcp.frankki.app/.well-known/oauth-authorization-server (OAuth 2.1, dynamic client registration) or a static bearer API key
  • Machine-readable overview: https://frankki.app/llms.txt

TypeScript

import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";

const transport = new StreamableHTTPClientTransport(
  new URL("https://mcp.frankki.app"),
  { requestInit: { headers: { Authorization: `Bearer ${process.env.FRANKKI_API_KEY}` } } }
);

const client = new Client({ name: "my-agent", version: "1.0.0" });
await client.connect(transport);

const { tools } = await client.listTools();
console.log(tools.map((t) => t.name));

const price = await client.callTool({
  name: "shipping_quote",
  arguments: { recipientCountry: "US", deliveryType: "standard" },
});

Python

from mcp import ClientSession
from mcp.client.streamable_http import streamablehttp_client

url = "https://mcp.frankki.app"
headers = {"Authorization": f"Bearer {FRANKKI_API_KEY}"}

async with streamablehttp_client(url, headers=headers) as (read, write, _):
    async with ClientSession(read, write) as session:
        await session.initialize()
        tools = await session.list_tools()
        print([t.name for t in tools.tools])

Authentication

Two modes, pick per client:

  • OAuth 2.1 for interactive clients (Claude Desktop, IDEs). The consent screen shows what the agent is allowed to do, and read-only scopes are granted by default. Best when an end user connects their own FrankKi account.
  • API key (bearer token) for headless agents, servers, and CI. Issue and rotate keys in the dashboard. Each key can carry a spending limit and an allowed-tools scope, so a key that may look up prices does not have to be a key that may spend money.

Wallet and pricing

FrankKi runs on a prepaid wallet. You top it up (via Stripe), and each send draws from the balance. No surprise post-paid invoices, and an out-of-credit agent simply cannot spend.

  • Pricing depends on the destination country. A domestic letter within Germany starts at 3,49 EUR; international sends are priced per destination.
  • Registered mail and other service levels are available where the destination supports them.
  • Volume tiers (Staffelpreise) reduce the per-letter price at scale.

Always call shipping_quote for the live, exact figure before order_send. Prices are authoritative from the server, never hardcode them.

Safety and human-in-the-loop

Sending physical mail is irreversible and costs money, so the platform is cautious by default:

  • Honest annotations. order_send and order_send_batch are annotated as destructive; read tools as read-only. Your agent can gate on this, and order_send accepts a maxCostEuros cap that aborts the send if the live price exceeds it.
  • Approval queue. Route sensitive letters through approval_submit for a four-eyes human check before anything is printed. Every decision is logged.
  • Per-key spending limits and tool scopes. Cap what an automated key can spend, and restrict which tools it may call.
  • Preview before commit. letter_create_draft returns a rendered preview and shipping_quote returns the cost, so an agent or a human can confirm before order_send.

Sandbox

Test the full flow without mailing anything. Request a sandbox key from the dashboard, point at the same endpoint, and you get a funded test wallet. Composition, pricing, and status all behave like production, but no paper is printed and no charge is made. Build and demo your integration end to end, then swap in a live key.

Compliance

Enterprise and EU-grade, for teams that need it:

  • GoBD-compliant archiving of every send, with retrievable receipts.
  • DATEV and generic DMS exports for accounting and document management.
  • DSGVO / GDPR: letter data is processed in the EU.
  • Platform and messages available in English and German.

How a send actually works

  1. Your agent calls letter_create_draft with the text and a recipient. FrankKi renders a professionally formatted document.
  2. shipping_quote returns the exact cost for that destination. The agent, or a human via the approval queue, confirms.
  3. order_send charges the prepaid wallet atomically and hands the job to fulfillment.
  4. FrankKi prints the letter and hands it off for delivery worldwide via our postal partners.
  5. Status flows back through order_status: printed, posted, and, where supported, delivered. order_einlieferungsbeleg returns the posting receipt.

Discovery and server manifest

For directories and registries, this repo publishes server.json (Official MCP Registry schema, name app.frankki/letters) and a repo-level llms.txt for AI-assisted discovery.

Found this server through an MCP directory? The canonical source of truth is always frankki.app/mcp.

Documentation

Doc Contents
docs/tools.md Every tool with scope, side effects, the canonical send flow, and error codes
docs/authentication.md OAuth 2.1, API keys, the full scope list, recommended setups
docs/clients.md Setup for Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Zed, OpenAI Agents SDK, LangChain
docs/use-cases.md Dunning, terminations, compliance mail, onboarding, agent-native products
examples/ Runnable TypeScript and Python clients (sandbox-first, send behind an explicit flag)

FAQ

Is the letter really printed and mailed, or is it email? Real paper. It is printed, franked, and delivered by postal partners. This is not email.

Which countries can it deliver to? Worldwide. Use shipping_quote to confirm reach and cost for a specific destination.

Do I host the server? No. FrankKi hosts it. This repo is documentation, server.json, and examples. You connect a client or agent to the hosted endpoint.

Is this for consumers or businesses? Primarily B2B and developer use: agents, backends, and automated workflows that need to dispatch physical mail at scale.

What if my agent goes rogue? It cannot spend beyond the wallet balance or a key's spending limit, sensitive letters can be forced through a human approval queue, and destructive tools are clearly annotated. Start in the sandbox.

Is there an API without MCP? Yes, FrankKi also exposes a REST API. The MCP server is the agent-native front door. See frankki.app/mcp.

Links

  • Homepage: https://frankki.app
  • MCP for humans and agents: https://frankki.app/mcp · https://frankki.app/en/mcp
  • Machine-readable overview: https://frankki.app/llms.txt
  • Endpoint: https://mcp.frankki.app
  • Registry name: app.frankki/letters (published in the Official MCP Registry, version 1.0.0)

Built by the team behind FrankKi: physical mail infrastructure for software, so your agents and workflows can send real letters worldwide without ever touching a printer.

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选