1,007 live data tools.
Loaded on demand.
1.7M+ requests/month 13K+ unique visitors/month
One MCP gateway URL gives your AI agent on-demand access to 295 sources — SEC, FDA, FRED, ClinicalTrials, USPTO, EPA, and ~290 more. Your agent loads only the ~20 tools it needs per session, so you don't burn context on definitions you'll never use. No stale training data. No rebuilding 50 integrations. No API key management.
npx pipeworx use weatherOr add any pack directly to Claude Desktop — no CLI required. Browse packs →
Why Pipeworx
Connector libraries built three billion-dollar businesses for the previous generation of tooling. We're doing it for AI agents — with the protocol, reliability, and edge-native architecture the agent era actually needs.
The connector library is the moat.
Airbyte ($1.8B, ~600 connectors) and Fivetran ($5.6B, ~700+) built billion-dollar businesses on one insight: developers don't want to write integrations one at a time. Plaid ($13.4B) proved the same model in a single vertical.
Pipeworx already has parity on connector breadth — 1,007 tools across 295 sources — and we built it in six weeks with Claude Code. The integration economics have changed; we're targeting the next consumer: agents, not data pipelines.
Full MCP, not just tool calling.
The MCP spec has three primitives — Tools, Resources, Prompts. Most MCP servers ship
only Tools. Resources let agents cite by URI for durable output
(pipeworx://edgar/company/{cik}/filings). Prompts encode optimal
multi-step workflows.
Pipeworx ships all three, plus five gateway-native meta-tools
(ask_pipeworx, discover_tools, resolve_entity,
compare_entities, remember/recall/forget)
that collapse 5–10 agent round-trips into one call.
Reliability is the product.
Agent tool calls fail 25–30% of the time in production against typical MCP servers. Almost no one measures it. We instrument every call, classify errors, and ship telemetry-driven fixes.
Last week we cut our own error rate from 27.6% to 6.9% in one afternoon — every remaining error is a legitimate user/config issue, not a bug. Read how →
vs Composio / Pipedream MCP / Zapier MCP: we ship the full MCP protocol surface and target autonomous-agent traffic, not human-triggered workflows. vs Smithery / mcp.so / Glama: they're directories; we're one MCP server with 295 sources behind it. vs DIY: every team rebuilds the same integrations one source at a time.
Featured packs
Twelve of 276 hosted packs · 1011 tools across the catalog · free tier · browse all 276 packs →
Edgar
EDGAR MCP — SEC EDGAR public APIs (free, no auth)
Fred
FRED MCP — Federal Reserve Economic Data (St. Louis Fed)
Openfda
OpenFDA MCP — wraps the openFDA API (free, no auth required)
Clinicaltrials
ClinicalTrials MCP — wraps ClinicalTrials.gov API v2 (free, no auth)
Attom
ATTOM MCP — Premium real estate data from ATTOM Data Solutions
Census
Census MCP — U.S. Census Bureau housing-relevant APIs.
Patents
Search US patents by keyword, retrieve patent details, and look up inventors via PatentsView
Epa Echo
EPA ECHO MCP — wraps EPA ECHO Web Services (free, no auth)
Samgov
SAM.gov MCP — Federal contract opportunities and entity registration data
Weather
Real-time weather conditions and multi-day forecasts from Open-Meteo
Wikipedia
Search and retrieve Wikipedia articles — summaries, section structure, and random article discovery
Crossref
Search and retrieve academic paper metadata — DOIs, authors, citations, and journals via Crossref
How it works
A real example: agent connects to Pipeworx, asks for Apple's recent SEC filings, gets structured data with citation URIs. One round trip.
Add Pipeworx to your MCP client
Drop one JSON block into claude_desktop_config.json, Cursor settings, or
any MCP-aware client. No API keys, no SDK install.
{
"mcpServers": {
"pipeworx": {
"url": "https://gateway.pipeworx.io/mcp"
}
}
}Your agent makes a real call
User asks a question; the model picks the right Pipeworx tool from the ~20 surfaced for this session and calls it directly. No fan-out, no orchestration code — just one JSON-RPC call to the gateway.
User: What did Apple file with the SEC in the last quarter?↓Agent calls: edgar_company_filings({ ticker_or_cik: "AAPL", limit: 5 })
Pipeworx returns structured data with citations
Every entity in the response carries a pipeworx:// URI — your agent can
quote these directly and downstream consumers can dereference them via MCP Resources.
Live data; SEC source-of-truth; no hallucination surface.
{
"cik": "0000320193",
"company_name": "Apple Inc.",
"filings": [
{
"accession_number": "0000320193-25-000123",
"filing_date": "2026-02-01",
"form": "10-Q",
"primary_document": "aapl-20260128.htm",
"document_url": "https://www.sec.gov/Archives/edgar/data/320193/...",
"_pipeworx_uri": "pipeworx://edgar/company/0000320193/filings/0000320193-25-000123"
},
{ "form": "8-K", "filing_date": "2026-01-30", "_pipeworx_uri": "pipeworx://edgar/..." },
{ "form": "8-K", "filing_date": "2026-01-15", "_pipeworx_uri": "pipeworx://edgar/..." }
],
"_meta": {
"tier": "anonymous",
"cached": false,
"latency_ms": 412,
"tool": "edgar.edgar_company_filings"
}
}Get started
Use the SDK in any TypeScript/Node project, or connect any MCP client — Claude Desktop, Cursor, Windsurf, or your own app.
import { Pipeworx } from '@pipeworx/sdk';
const px = new Pipeworx();
const result = await px.ask("What is the US trade deficit with China?");{ "mcpServers": { "pipeworx": { "url": "https://gateway.pipeworx.io/mcp" } } }