Ask your agent what a company actually sells.

Same job as the API: a domain in, cited standalone products out. Sales uses it before outreach. Intel uses it to watch competitors. Agents use it inside Cursor or Claude. Remote MCP over HTTP — nothing to install, no npm package, no local process to keep alive.

No package to install or update
No local process — the server is always current
Works with any MCP client that supports HTTP transport

Add to Cursor with your key or copy a block

~/.cursor/mcp.json

{
  "mcpServers": {
    "stacksift": {
      "url": "https://stacksift.in/mcp-server/mcp",
      "headers": {
        "Authorization": "Bearer sk_live_YOUR_KEY"
      }
    }
  }
}

Replace YOUR_KEY with a real key from /app/keys. Minting a key there also gives you a one-click Add to Cursor with the token already in it. Placeholder install: cursor:// with YOUR_KEY — that 401s until you swap it.

Claude Code

claude mcp add --transport http stacksift \
  https://stacksift.in/mcp-server/mcp \
  --header "Authorization: Bearer sk_live_YOUR_KEY"

Then paste a prompt from below. analyze_domain takes 1–2 minutes.

.vscode/mcp.json

{
  "servers": {
    "stacksift": {
      "url": "https://stacksift.in/mcp-server/mcp",
      "headers": {
        "Authorization": "Bearer sk_live_YOUR_KEY"
      }
    }
  }
}

VS Code uses servers, not mcpServers. Same URL and Bearer header as Cursor.

Custom GPT · Actions

1. Get a key: https://stacksift.in/app/signup
2. Custom GPT → Create → Actions → Import from URL:
   https://stacksift.in/openapi.json
3. Authentication: API Key. Auth type: Bearer. Paste sk_live_…
4. Instructions (paste):
You call Stacksift to list standalone purchasable products for a B2B domain.
POST /v1/analyze with {"domain":"example.com"}. Timeout 180s.
Report only verdict=true_product. Quote evidence_urls. Never invent URLs.
If requires_human_review is true, say so. Ignore features, tiers, and add-ons.

Same meter as the API. Import only /openapi.json — that spec is the customer API, not admin.

Most MCP servers ask you to run npx -y some-package on your machine. Stacksift doesn't need to.

Agent instruction · paste once

You have Stacksift tools. When asked what a company sells:
1. Call analyze_domain with a bare domain (no https, no path).
2. List only products with verdict true_product.
3. Quote at least one evidence_urls item per product. Never invent URLs.
4. Mention requires_human_review if it is set.
5. Ignore features, pricing tiers, and add-ons.
analyze_domain is synchronous (about 1–2 minutes). Failed calls are not charged.

Four steps, one email

1You're here.

2Enter your email below.

3Click the link — you're in. Mint a key on /app/keys; that page shows the config block with the key in it, once.

4Copy the block. Ask your agent what a prospect actually sells.

50 analyses / month · no card · key shown once, stored as a hash

01 · Seven tools

Exactly the ones the API has.

analyze_domain	One B2B domain → real products with cited evidence URLs. 1 analysis. Sync, typically 1–2 minutes.
analyze_batch	Up to 500 domains. Returns `job_id`. 1 analysis each.
get_job	Poll a batch started with `analyze_batch`.
cancel_job	Stop a running batch.
get_credits	Analyses left and when they expire.
get_usage	Recent calls for this key.
get_review_queue	Products flagged `requires_human_review`.

Input is a bare domain like hubspot.com · every call runs the live pipeline and uses 1 analysis · failed calls are not charged · same meter as POST /v1/analyze · batch max 500

02 · What to type

Five prompts. Copy, paste, run.

analyze_domain

Use Stacksift analyze_domain on hubspot.com. List only true_product SKUs with confidence and one evidence URL each. Ignore features and pricing tiers.

analyze_domain

What does gong.io actually sell? Call Stacksift and quote the evidence URLs. If anything is flagged requires_human_review, say so.

analyze_batch + get_job

Use Stacksift analyze_batch on ["zapier.com", "make.com", "n8n.io"]. If you get a job_id, poll get_job until done, then compare the product lists.

get_credits

Call Stacksift get_credits. Then analyze_domain on dialpad.com. Report analyses left before and after.

get_review_queue

Call Stacksift get_review_queue. If it's empty, analyze_domain on salesforce.com and check the queue again for anything a human should confirm.

03 · Fine print

Same API. POST /mcp-server/mcp is JSON-RPC in front of /v1 — same key, same limits, same meter.

No key? The server returns 401. Get a key from this page, then paste the block.

No analyses left? Tools return JSON with buy_url the model can relay, not a silent failure.

Slow analysis? analyze_domain is synchronous (about 1–2 minutes). Batch jobs: poll get_job. Repeats run the pipeline again.