For AI agents · MCP

Give your agent the Stackcut catalog

Any MCP client can search 63 replacement recipes, re-cost alternatives against a real bill, pick the cheapest stack for a new project at its real volume, read your own Stackcut savings plan with a read-only code, and pull a build spec for a coding agent. Free, no API key.

https://stackcut.io/mcp

Transport: Streamable HTTP
Auth: none
Tools: 12
Plain JSON: /data/recipes.json

Connect

Claude Code

claude mcp add --transport http stackcut https://stackcut.io/mcp

Claude (web and desktop)

Settings → Connectors → Add custom connector. Name it Stackcut and paste the URL. No sign-in step.

Cursor

Add to ~/.cursor/mcp.json

{
  "mcpServers": {
    "stackcut": {
      "url": "https://stackcut.io/mcp"
    }
  }
}

VS Code

Add to .vscode/mcp.json

{
  "servers": {
    "stackcut": {
      "type": "http",
      "url": "https://stackcut.io/mcp"
    }
  }
}

Any client or a quick test

Stateless JSON-RPC over POST. Clients that only speak stdio can use npx mcp-remote https://stackcut.io/mcp.

curl -s https://stackcut.io/mcp -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"find_alternatives","arguments":{"vendor":"Airtable","monthly_spend_usd":400}}}'

Your savings plan, in your agent

Got a Stackcut plan? Hand it to your agent and let it do the legwork: compare plans, draft cancellations, set up replacements. It reads the plan with a read-only code that expires, and asks you before any cancel or purchase.

1. Make a code

Open your plan on stackcut.io/app → Send to my agent → Make a code. It lasts 7 days; “Stop sharing” turns it off.

2. Paste the prompt

Copy the prompt it gives you into Claude, ChatGPT, Cursor or Claude Code. Your agent calls:

get_my_plan({"code": "sca_…"})

What your agent gets

Every line with its verdict and proof: the receipts or card charges behind the price, the alternative's price with a link to its pricing page, which of your ticked features it covers, the steps, checks, pros and cons, time, and a ready prompt. A saving that can't be proven is null, never a guess. No MCP in your agent? “Copy plan as Markdown” on the same card.

Choosing a stack

When you ask your agent “what's the best database for this?” or “what's the cheapest way to build this?”, choose_stack compares vendors per job at your volume. A tier counts as a fit only when its published limit covers your number; otherwise it says what to check. Open-source options are listed with their license, and your server cost is left to you, not guessed.

choose_stack({"project": "B2B SaaS MVP", "needs": [
  {"job": "database", "features": ["Postgres database", "Auth"], "volume": {"amount": 20000, "unit": "MAU"}},
  {"job": "email_transactional", "volume": {"amount": 40000, "unit": "emails_per_month"}},
  {"job": "hosting"} ]})

→ Supabase Pro $25/mo (“100k MAU” covers 20,000; the free tier may also fit but doesn't state MAU), SendGrid Essentials $19.95/mo (“50,000 emails/mo” covers 40,000), and for hosting a question: how much traffic?

Audit a codebase

Your agent can see what your code actually uses: payment, email, analytics, monitoring and AI SDKs, env var names, config files and API calls. The Stackcut skill scans the repo on your machine and sends only vendor and feature names to audit_stack. No source code, no secret values, and it never opens .env.

1. Install the skill

Download it and unzip into your agent's skills folder. For Claude Code: ~/.claude/skills/ (so you get ~/.claude/skills/stackcut/SKILL.md).

Download the Stackcut skill

2. Ask your agent

"Audit the paid services in this repo with Stackcut." It runs the scanner, asks what you pay, then calls audit_stack.

node ~/.claude/skills/stackcut/scripts/detect.mjs . --latest --pretty

What the scan sends

Only a summary like this. Rules for about 100 vendors live at /data/detect.json (also the get_detection_rules tool).

{"vendor":"sendgrid","features_used":["Transactional send","Dynamic templates"],
 "evidence":["npm:@sendgrid/mail","env:SENDGRID_API_KEY"],"monthly_spend_usd":19.95}

Tools

Tool	What it does	Details
search_recipes	Search replacement recipes	Search Stackcut's catalog of replacement recipes.
get_recipe	Get a recipe	Full recipe: the features it covers, what is out of scope, checks to run before switching, and every replacement path with modeled first-year net savings, monthly costs, setup hours, maintenance, pros/cons, license notes and sources.
find_alternatives	Find alternatives for a subscription	Given a product the user pays for (and optionally what they pay per month and the features they use), return the ranked ways to replace it, re-costed against their actual monthly bill, with which of their features each path covers.
get_build_packet	Get a build packet (Replace with AI)	Agent-ready build packet for a recipe: SPEC.md (must-haves, out of scope, data model, checks, cost targets, acceptance tests, deliverables), AGENTS.md and CLAUDE.md instructions, and recipe-manifest.json.
leaderboard	Savings leaderboard	Top replacement paths across the catalog: 'savings' ranks by modeled first-year net savings, 'payback' by months to recover setup cost, 'popular' by how often people and agents used them.
audit_stack	Audit a stack (code + billing)	Audit the paid services a codebase or team uses.
get_detection_rules	Get code detection rules	Rules for finding paid services in a codebase locally: package names (npm, PyPI, RubyGems, Go), environment variable NAME patterns, config files, API hosts, and per-feature code patterns for about 100 vendors.
list_categories	List categories	Categories in the catalog with recipe counts and the best modeled first-year savings in each.
get_my_plan	Get the user's Stackcut savings plan	Read the user's own Stackcut savings plan with the read-only agent code they copied from stackcut.io (Send to my agent; it starts with sca_).
choose_stack	Choose the most cost-effective stack	For a project being built or planned, compare vendors per job (database, auth, hosting, transactional email, analytics, error monitoring, storage, payments, LLM API, …) at the user's expected volume.
list_stack_jobs	List stack jobs	The jobs choose_stack can compare (database, auth, hosting, email, …) with how many vendors and open-source options each has.
report_outcome	Report what the user did	Anonymously count that a user picked, switched to, or reverted from a replacement path.

Try asking your agent

"Audit the paid services in this repo and tell me what to cut."

Skill scan → audit_stack → per service: eligible, needs your price, keep, or no recipe yet.

"We pay $1,000 a month for Salesforce Pro and only use contacts, one pipeline and tasks. What should we do?"

find_alternatives → Bigin Express, modeled at $10,260 saved in year one, with what it does not cover.

"Which open-source tools could replace our Semrush seat for keyword and rank checks?"

search_recipes + get_recipe → OpenSEO self-hosted or hosted, with setup hours and license notes.

"Build us a replacement for our Typeform intake flow."

get_build_packet → SPEC.md, AGENTS.md and a manifest your coding agent builds from.

"What are the biggest savings for a small team right now?"

leaderboard → the same ranking as stackcut.io/leaderboard.

"Here's my Stackcut code. Help me cancel what I don't need."

get_my_plan → your plan line by line with its proof, then the steps, asking you first.

"Figure out the most cost-effective way to build this app for 5,000 users."

choose_stack → per job: the cheapest tier whose published limits cover 5,000 users, with links.

What the server keeps

No accounts, and the server stores no user data. get_my_plan passes your plan through to your agent with your read-only code and keeps nothing. Each recipe view, estimate, spec download and reported switch adds 1 to a public daily counter for that recipe (that's the "Most used" leaderboard). Your IP address is used only as a rate-limit key, hashed with the date and deleted after 2 days. Savings are modeled, so agents are told to confirm must-have features before recommending a switch.