AI agents & MCP

Overview

• MCP server at https://depositscout.com/api/mcp: every Data API endpoint as a tool, for Claude, Cursor or any MCP client. It uses your API key and costs exactly what the REST API costs.
• WebMCP: agents running in a visitor's browser get a few free, read-only tools that show a short sample of our pages.
• For crawlers and assistants: /llms.txt, Markdown versions of every page, an API catalog and agent skills. See discovery files.

MCP server

The server speaks MCP over Streamable HTTP at https://depositscout.com/api/mcp. It's stateless, so any MCP client that supports remote HTTP servers can connect. Send your API key as a header: ds_test_… for free sample data, ds_live_… for real data.

Listing tools needs no key. Every tool call goes through the same gateway as the REST API: the same scopes, plans, rate limits, IP allow-list, credits and usage log. A live key is charged exactly as the matching endpoint is, so lists cost per row returned and a small limit keeps the cost down. Calls made through MCP show on your Usage page like any other request.

A tool returns the endpoint's normal response, { data, meta } with meta.creditsCharged and meta.creditsBalance, as both text and structured content, plus a disclaimer field (information, not financial advice) that your agent should pass on. An API error (no key, not enough credits, wrong plan) comes back as a tool error with the usual error body, and isn't charged. Start with the free ping and me tools to check the key, your balance and which tools it can call.

Connect a client

claude mcp add --transport http depositscout https://depositscout.com/api/mcp \
  --header "Authorization: Bearer ds_test_YOUR_KEY"

Cursor reads .cursor/mcp.json; Claude Desktop and most other clients take the same url and headers. The server card is at /.well-known/mcp/server-card.json.

Tools

One tool per endpoint. A tool's inputs are its endpoint's parameters, and the endpoint's docs page has the details. Send your key with tools/list and you get only the tools that key can call (its scopes, plan and mode), so an agent never picks one it would be refused on. Every tool's _meta has its price in machine-readable form, "com.depositscout/cost": { "credits": 1, "unit": "row" }, for showing or budgeting spend.

Prompts

Ready-made workflows your MCP client can offer its user. A prompt is only a recipe, so it's free; the tool calls it leads to are charged as usual. Each one tells the model to explain with the tools' own facts (matchReasons and the simulation's assumptions), keep lists small, and say it isn't financial advice.

• compare_savings (type, amount): Search, compare and simulate the top accounts of one type for a balance.
• check_savings_rate (sourceKey, amount): Look up one account, simulate it (with and without a rate cut) and see how it compares.
• explain_rate_history (sourceKey): Lay out how one account's rate has changed, with dates, without speculating why.

Resources

Free context an agent can read to explain the numbers, with no key and no credits: depositscout://glossary (account types and terms), depositscout://disclaimer, the Data Licence, FSCS protection, how we rank accounts, the guides index, and any guide through the https://depositscout.com/guides/{slug} template. Pages come back as Markdown, with rate tables shortened like the Markdown pages. The prompts' type and sourceKey arguments support completion, so a client can offer real values instead of guesses.

Build an agent

Connect with the official MCP SDK, give your model the tool list, and run the tool calls it asks for. Two tools help an agent answer without guessing: every rates item carries matchReasons (why it's in the result, as facts), and simulate works out what a balance would earn, including an intro bonus ending or a rate change.

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

const client = new Client({ name: "my-savings-agent", version: "1.0.0" });
await client.connect(
  new StreamableHTTPClientTransport(new URL("https://depositscout.com/api/mcp"), {
    requestInit: { headers: { Authorization: `Bearer ${process.env.DS_API_KEY}` } },
  }),
);

// Hand these to your model as tools; run the calls it asks for.
const { tools } = await client.listTools();

const rates = await client.callTool({ name: "rates", arguments: { type: "easy-access", amount: 50000, limit: 3 } });
const top = (rates.structuredContent as { data: any[] }).data[0];
const sim = await client.callTool({ name: "simulate", arguments: { sourceKey: top.sourceKey, amount: 50000 } });
console.log(top.accountName, top.matchReasons, (sim.structuredContent as { data: any }).data.result);

// Or start from a ready-made workflow.
const prompt = await client.getPrompt({ name: "compare_savings", arguments: { type: "cash-isa", amount: "20000" } });
await client.close();

Use a ds_test_… key while you build: every tool returns sample data for free, in the same shape as live data.

WebMCP

When an AI agent runs inside a visitor's browser (for example Chrome with WebMCP), every depositscout.com page offers it these read-only tools:

• find_savings_rates: the top rates for one account type (easy-access, fixed-rate, notice, cash-isa and more)
• look_up_bank: one bank or building society's best rates, accounts and switch offers
• list_switch_offers: current account switching bonuses
• read_page: any public page, such as /guides or /fscs-protection

They're free and need no key, and they return the same shortened Markdown as the pages: each table and list is cut to its first 3 rows, with a link to the full page. For complete, structured data, use the MCP server or the REST API.

Discovery files

• /llms.txt: what DepositScout is and its main pages
• Any page with Accept: text/markdown: that page as Markdown (tables shortened, as above)
• /.well-known/api-catalog: the API's OpenAPI, docs and status (RFC 9727)
• /.well-known/mcp/server-card.json: this MCP server
• /.well-known/agent-skills/index.json: skills for reading the site and calling the API
• /.well-known/ai-catalog.json: all of the above in one manifest
• Every page sends a Link header pointing at these.

Using the content

Search engines and AI assistants are welcome to read our pages and cite them in answers; please link to the page. We don't allow our content to be used to train AI models (Content-Signal: ai-train=no in robots.txt). Data from the API and the MCP server is covered by the Data Licence.