Kordi — Subscription Intelligence

Kordi MCP Server

Track, analyze, and act on your streaming & SaaS subscriptions — from any AI agent.

Kordi is a subscription intelligence platform. This Model Context Protocol server lets any MCP-speaking AI client — Claude Desktop, Cursor, Raycast, or your own agent — read a user's tracked subscriptions, score their billing health, surface savings, generate cancel/pause links, and push newly discovered subscriptions in.

It's a hosted, remote MCP server. There's nothing to install or run locally — you connect to a URL with a token.

• Endpoint: https://kordiapp.com/mcp
• Transport: Streamable HTTP
• Auth: Authorization: Bearer
• Registry: com.kordiapp/kordi-mcp-server on the official MCP Registry
• Docs: kordiapp.com/developers

Quick start

1. Get a token

Sign in at [kordiapp.com](https://kordiapp.com) and generate an API token at [kordiapp.com/token](https://kordiapp.com/token?ref=github). Tokens are valid for 90 days and grant read/write access to *your own* subscriptions only.

2. Connect your client

Claude Desktop — add this to your claude_desktop_config.json (see examples/):

{

"mcpServers": { "kordi": { "type": "streamable-http", "url": "https://kordiapp.com/mcp", "headers": { "Authorization": "Bearer YOUR_TOKEN_HERE" } } } }

Any client / raw HTTP — pass the token as a Bearer header (preferred — it never lands in logs, Referer, or history):

# List the available tools

curl -X POST "https://kordiapp.com/mcp" \ -H "Authorization: Bearer YOUR_TOKEN_HERE" \ -H "Content-Type: application/json" \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

A legacy ?token= query-string form is also accepted for back-compat, but prefer the header.

3. Ask away

*"What am I paying for each month?"*

*"Which subscriptions could I pause to save money?"*

*"Do I have any price hikes or trials ending soon?"*

*"I just signed up for Apple TV+ at $9.99/mo — track it."*

Tools

Plus a resource:

A platform-analytics tool (kordi_get_analytics) also exists but is restricted to administrators.

Tool detail

kordi_list_subscriptions — read

Args

• include_paused (boolean, default false) — include paused subscriptions
• offset (number, default 0) — pagination start
• limit (number, 1–100, default 50) — page size

Returns

{

"subscriptions": [ { "id": "netflix", "name": "Netflix", "price": 15.99, "billingDate": 12, "state": "ACTIVE" } ], "count": 1, "total_count": 1, "offset": 0, "limit": 50, "has_more": false, "total_monthly": 15.99 }

</details>

kordi_analyze_billing_health — read

Takes no arguments. Returns total monthly spend, active/paused counts, price_hikes, unused_services, savings_potential, top_savings_candidate, billing_clusters, and a health_score (0–100, higher is healthier).

kordi_get_cancellation_route — action

Args

• service_id (string, required) — the exact id from kordi_list_subscriptions
• reason (too_expensive | not_using | switching_service | finished_content, optional)

Returns { "success": true, "service": "Netflix", "cancellation_url": "https://…" }. Call only when the user explicitly asks to cancel or pause.

kordi_ingest_subscription — write

Args

• name (string, required) — e.g. "Disney+"
• amount (number, required) — monthly amount in USD
• bill_date (number 1–31, or ISO date string) — billing day or date
• source (qira | screenpipe | manual, optional) — discovery-source attribution

Dedupes by name (re-calling updates amount/bill date). Returns { status, subscription, pause_link, dashboard_url }.

kordi_get_top_shows — read

Args

• limit (number, 1–50, default 10)
• sort_by (swipes | likes | like_rate, default swipes)

Returns a cross-user popularity leaderboard. Aggregate only — no per-user data.

Auth & identity model

• Your token resolves server-side to your Kordi account. Every tool call is scoped to your account — you can only ever see and act on your own subscriptions.
• Identity is bound once per session at the edge; a session cannot be re-bound to a different user mid-stream.
• Sensitive card fields (PAN, CVV, expiry, raw token) are stripped from every response — they never leave the server.
• Tokens are short-to-medium-lived (90 days). Treat them as secrets. You can rotate at kordiapp.com/token at any time.

Limits

• Rate limit: 20 tool calls per 60 seconds per session. Over the limit, tools return a back-off message instead of hitting the database — your agent should wait, not retry-loop.
• Response size: large list responses are capped (~25k chars); page with offset/limit for big accounts.
• Pause/cancel links are signed and expire after 7 days.

For integration partners

Discovery sources (email-intelligence agents, on-screen detection pipes, etc.) can push subscriptions into a user's Kordi account via kordi_ingest_subscription, or provision brand-new users via the public POST /api/guest-ingest endpoint. See kordiapp.com/developers for the partner provisioning flow, source attribution, and the x-partner-key model.

Links

• Web app: kordiapp.com
• Developer docs: kordiapp.com/developers
• Get a token: kordiapp.com/token
• MCP Registry listing: com.kordiapp/kordi-mcp-server
• Model Context Protocol: modelcontextprotocol.io

License

Documentation and the registry manifest in this repository are released under the MIT License. The Kordi server, brand, and backend are proprietary; this repo does not contain server source code.