One URL. Every scraper. Your AI agent finds them all.
The Stekpad MCP Gateway introspects every robot in your workspace and exposes each one as a named tool. Claude, Cursor, Continue — they discover your data without a single prompt-engineered API reference.
Most “AI scrapers” were built for humans clicking buttons.
An agent hitting a REST API with a random key and a README in its context window isn't discovering anything. It's pattern-matching over prose. MCP is what tool discovery looks like when it's a native contract.
The Model Context Protocol is an open standard for letting an AI client list, inspect, and call tools at runtime — no hardcoded schema, no vibe-coded API wrapper. The agent asks the server “what can you do?”, gets back typed, named tools, and calls them.
Stekpad is built for that. A scraper produces lists. Lists update. Agents want to query “what changed since Tuesday?” That's exactly the shape MCP serves well. So we wired it end-to-end.
The result: one URL per workspace. Every robot you create becomes a tool your agent discovers the next time it reconnects. You name a robot YC W26 companies and Claude gets get_yc_w26_companies() for free. No config. No drift. No hallucinated endpoints.
- ● get_<slug>(limit?, cursor?)
- ● run_<slug>(dry_run?)
- ● diff_<slug>(since)
- ● describe_<slug>()
- ● search_<slug>(query)
Plus workspace-level tools — list_robots(), get_quota(), create_robot(url) — so your agent can bootstrap new scrapers on its own.
Paste it once. Every robot is live.
{ "mcpServers": { "stekpad": { "url": "https://mcp.stekpad.com/stkpd_live_9a8b7c6d" } }}Restart Claude Desktop once. Your tools appear in the menu.
{ "mcp.servers": { "stekpad": { "url": "https://mcp.stekpad.com/stkpd_live_9a8b7c6d" } }}Same URL works across every MCP-compatible client. No polling, no refresh — the gateway streams tool changes.
What your agent sees.
You → Use stekpad. What's new in YC W26 companies this week?
Claude → calling diff_yc_batch_w26(since="7d")
→ 14 new companies added, 2 removed, 8 fields changed
Here are the 14 newcomers, grouped by batch track:
• AI/Agent (6): Corridor, Alloy, Tangent, …
• Fintech (3): Vault, Piper, Ledgerly
• Healthtech (2): Cadence, Noa
• Hardware (3): Gantry, Muon, Stipple
Two removals: Acme.ai and ForgeOS (batch withdrew).
Want me to pull the full profile for any of these?No prompt about an API key. No "here's how my endpoint works". The tool was named, typed, and discovered — the rest is a normal conversation.
Dynamic introspection. Zero config drift.
Tools generated from robot slugs
Every robot in your workspace gets its slug normalized (lowercase, underscored) and mapped to a tool name. Rename a robot, the gateway re-emits — your agent reconnects and sees the new name.
One key, per-workspace scoping
The URL carries your workspace key. We validate on every MCP handshake, enforce RBAC, and rate-limit per robot. Revoke the key in the dashboard and every agent disconnects cleanly.
Tool list streams over SSE
When you create a new robot or archive one, the gateway pushes a tools/list_changed event. MCP clients that support it refresh automatically. No reload, no restart.
Per-tool quota and audit
Every tool call logs workspace, tool name, latency, rows returned, and cost. Export the audit log as JSON or stream it to your SIEM. Budget per robot in the dashboard.
Works with the clients you already use.
Give your agent real data.
Your MCP URL is ready the moment you sign up. First robot in 30 seconds. First MCP call in 60.