Magellan Data Spotlight

Magellan Data MCP

Corporate ownership intelligence for AI agents. Connect Claude, Cursor, or any MCP-compatible client to Magellan Data's Spotlight platform and ask who owns whom — parent companies, PE backing, corporate families, and portfolio siblings — over lists of thousands of companies at a time.

Hosted server. No install. Connect with an API key, or via OAuth from hosted connectors like claude.ai.

https://spotlight-mcp.magellandata.io/mcp

That's an MCP endpoint, not a web page — paste it into your client's config, not your browser. Opening it directly returns an error, which is the correct behaviour for every hosted MCP server.

---

Try asking

Once connected, these all work in plain language:

"Read accounts.xlsx and tell me which of these companies are PE-backed, and by whom."

"We just closed Moosend. Find every other company owned by the same PE firm — those are

our warm expansion targets."

"Which of my 200 accounts are subsidiaries of a larger parent? Append the parent company

to the spreadsheet and save it."

"Map the full corporate family under Constant Contact's parent so I can see the sister

companies I'm not covering."

The agent handles the whole pipeline — submitting the batch, polling until the runs finish, checking the price, and joining results back to your original file by input_url.

---

Quick start

1. Get an API key. Sign up at magellandata.io and generate a key from the Spotlight dashboard. Keys look like mgln_sk_live_…. Every key carries its own credit balance, and usage is billed to the key that made the call.

2. Connect your client.

Claude Code

claude mcp add --transport http magellan-data https://spotlight-mcp.magellandata.io/mcp \

--header "Authorization: Bearer mgln_sk_live_YOUR_KEY"

claude mcp list # expect: magellan-data ... ✓ Connected

</details>

Cursor

Add to ~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{

"mcpServers": { "magellan-data": { "url": "https://spotlight-mcp.magellandata.io/mcp", "headers": { "Authorization": "Bearer mgln_sk_live_YOUR_KEY" } } } }

</details>

VS Code

Add to .vscode/mcp.json — this prompts for the key rather than storing it in the file:

{

"inputs": [ { "type": "promptString", "id": "magellanKey", "description": "Magellan Data API key (mgln_sk_live_...)", "password": true } ], "servers": { "magellan-data": { "type": "http", "url": "https://spotlight-mcp.magellandata.io/mcp", "headers": { "Authorization": "Bearer ${input:magellanKey}" } } } }

</details>

Any other MCP client

The server speaks Streamable HTTP and is stateless. Point any compliant client at https://spotlight-mcp.magellandata.io/mcp with an Authorization: Bearer header.

3. Verify. Ask your agent: *"Check my Magellan Data credit balance."* If it comes back with a number, you're connected. That call is free.

---

Authentication

Two methods are supported; use whichever fits how you're connecting.

API keys (mgln_sk_live_…) — for local and CLI clients like Claude Code, Cursor, or anything you configure by hand. Generate a key in the Spotlight dashboard and pass it as a Bearer token in the Authorization header, as shown in the quick start above. Usage is billed to the key that made the call.

OAuth 2.1 — for hosted connectors such as claude.ai and ChatGPT. There's nothing OAuth-specific to set up on your side: when you add the connector (from the directory, or as a custom connector where the platform holds the client credentials), you'll be sent to Spotlight to sign in and approve access — the usual "connect, sign in, allow" flow — and the client handles tokens from there. Usage is billed to the Spotlight account you approved with.

Either way, every request is authenticated and attributed to one Spotlight account — the two methods differ only in how the credential reaches the server.

---

What you can look up

For parent_companies and pe_ownership you pay only for hits: a URL where nothing is discovered costs nothing, so 70 credits × your URL count is the *ceiling*, not the bill. The other two are priced per output record. Either way the exact price is known once a run completes — the agent reads it off the run and tells you before downloading.

---

How credits work

This is the part worth understanding, because it's what keeps an autonomous agent from quietly burning your balance:

• Submitting a batch is free. So is starting the processing runs, polling their status, and checking your balance.
• Credits are charged at exactly one point: download_run. Nothing before it costs anything.
• The agent surfaces the price first. Every run carries a price field, and the server instructs connected clients to state that number and wait for you before downloading.
• Re-downloads are free, forever. Once a run is purchased it stays purchased. If a download link expires, just ask again — no second charge.

If you'd rather not think about it: everything up to and including "how much will this cost" is free, and the agent will ask.

---

Tools

The flow is always submit → process → poll → download. Results come back as a short download link (https://api.magellandata.io/dl/…, valid for 1 hour) rather than inline, so a 2-million-row result set never touches your agent's context window. The link needs no auth header — hand it straight to curl, pandas.read_csv, or a browser. In hosted chat environments without general internet access (like the claude.ai connector), the assistant presents it as a clickable link for you to open rather than fetching it itself.

There's also a `magellan_playbook` prompt — a GTM strategy guide covering expansion- target ranking, deal_type interpretation, and MSA-timing triggers. In Claude Code it shows up as a slash command under the magellan-data server.

---

Working example

examples/quickstart.py runs the full pipeline against the REST API directly — useful for understanding what the agent is doing under the hood, or for wiring Magellan Data into a non-agentic pipeline.

export MAGELLAN_API_KEY=mgln_sk_live_YOUR_KEY

python examples/quickstart.py --urls moosend.com constantcontact.com --type pe_ownership

examples/accounts.csv and accounts.xlsx are a 12-row fixture you can point an agent at without supplying your own data. The companies span independents, acquired subsidiaries, and PE-backed businesses, so each output type returns something. Enriching all twelve costs at most 840 credits per match-based output type (only the companies where something is found are charged) — cheap enough to validate the pipeline for real.

It stops and asks before the download step, so you can run it end to end without spending credits until you mean to.

examples/PROMPTS.md has a walkthrough of agent prompts in increasing order of ambition, with the expected cost of each noted up front.

---

Skills

The skills/ directory holds task-focused playbooks that turn raw tool output into a finished deliverable — a ranked target list, a coverage map, an enriched spreadsheet — instead of JSON. Clients that support skills can load them from this repo; clients that don't can still be pointed at the markdown directly.

---

Repository contents

.mcp.json                     MCP registration (Claude Code, Codex)

mcp.json MCP registration (Cursor) .vscode/mcp.json MCP registration (VS Code, prompts for key) .claude-plugin/ plugin.json Claude plugin metadata marketplace.json Claude marketplace metadata server.json Official MCP Registry metadata skills/*/SKILL.md Task playbooks examples/ quickstart.py Full pipeline against the REST API PROMPTS.md Agent prompt walkthrough accounts.csv 12-row test fixture accounts.xlsx Same fixture, spreadsheet form assets/ logo.png Wordmark, navy (for light backgrounds) logo-light.png Wordmark, white (for dark backgrounds) social-preview.png 1280x640 social card — upload under Settings > Social preview social-preview-light.png Light-background alternate

To use the manifests locally:

git clone https://github.com/sorrek/mcp.git

Then point your client at the repo root or the specific manifest path.

---

Troubleshooting

The endpoint URL shows an error in my browser. Expected. /mcp speaks JSON-RPC over POST; the transport spec reserves GET on that path for opening an event stream, so a server like this one that doesn't offer streams answers with

1. There's nothing to see there. To check the server is actually up, ask your agent to

call check_credits — that's free and exercises the whole auth path.

My client connects but lists no tools. Almost always an auth problem rather than a transport one. Confirm the header is Authorization: Bearer mgln_sk_live_... with the Bearer prefix, and that the key is active in the Spotlight dashboard.

A download link says it has expired (HTTP 410). Download links last an hour. Call download_run again for a fresh one — re-downloads are free.

A run has been `processing` for a long time. Large batches take minutes, not seconds. Poll every 10–30 seconds rather than in a tight loop. If a run sits past 20 minutes, open an issue with the run_id.

---

Notes and limits

• Batch size: up to 5,000 URLs per batch. Larger lists should be chunked; the agent will do this if you ask.
• Timing: large runs take minutes, not seconds. Polling every 10–30 seconds is the right cadence and the server instructions say so.
• **Overlapping inputs: