← Home

Your Claude stack with Pipeworx

Pipeworx is the live-data layer of a Claude agent stack. It does not replace your filesystem, your repo, or your SaaS actions — but it handles ~290 live data sources, agent memory, and citation URIs behind one URL. Here is how it fits in a typical Claude Code / Claude Desktop / Cursor / VS Code stack, with honest pairing recommendations for every layer we do not cover.

Recommended baseline (three MCP servers)

A working stack for most agent tasks. Three servers, one data layer, one filesystem, one repo. Replace the GitHub server with GitLab if you prefer; everything else stays. 

{
  "mcpServers": {
    "pipeworx": {
      "url": "https://gateway.pipeworx.io/mcp"
    },
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/path/to/project"
      ]
    },
    "github": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-github"
      ],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_…"
      }
    }
  }
}

Add this to claude_desktop_config.json on Claude Desktop, the equivalent config on Cursor / VS Code / Windsurf, or via claude mcp add for Claude Code. The Pipeworx gateway works without an API key on the free tier (50 calls/day anonymous, 2,000/day with a free account).

The eight zones of a typical stack

An agent doing real work touches eight functional zones. Pipeworx covers three of them outright, partials a fourth, and pairs cleanly with established servers for the rest.

Live data

Covered by Pipeworx

Where do facts come from?

1,498 tools across 406 packs — SEC EDGAR, FDA, FRED, USPTO, EPA, Census, ClinicalTrials, ATTOM, UN Comtrade, and ~280 more. One gateway URL.

Memory

Covered by Pipeworx

How does the agent retain state across sessions?

Built-in remember / recall / forget meta-tools. State persists on the gateway, keyed to your account or session.

  • @modelcontextprotocol/server-memory (if you prefer a standalone)

Citations

Covered by Pipeworx

How does the agent reference what it just looked up?

Every tool returns stable pipeworx:// resource URIs (e.g. pipeworx://edgar/company/0000320193/filings). Resolvable via resources/read; safe to embed in agent output.

Files & local IO

Out of scope — pair with

How does the agent read or write files on your machine?

Out of scope for Pipeworx — the gateway is server-side and has no access to your local filesystem.

  • @modelcontextprotocol/server-filesystem (official, scope a directory)

Code & repos

Out of scope — pair with

How does the agent open PRs, read issues, or commit code?

Out of scope. Pipeworx tracks USPTO patents and SEC filings, not your repo.

  • @modelcontextprotocol/server-github (official, repo + issues + PRs)
  • @modelcontextprotocol/server-gitlab

SaaS actions

Out of scope — pair with

How does the agent send Slack messages, create Notion pages, or update CRM records?

Out of scope. Pipeworx is read-mostly live data, not write-side SaaS automation.

  • Composio (200+ SaaS apps)
  • Pipedream MCP (2,500+ integrations with triggers)
  • Zapier MCP (broadest SaaS coverage)

Browser automation

Out of scope — pair with

How does the agent navigate the web or scrape rendered pages?

Out of scope. Pipeworx wraps APIs, not headless browsers — when you need a real DOM, pair with a browser server.

  • @modelcontextprotocol/server-puppeteer
  • Browserbase MCP (managed headless)

Generic HTTP

Partial — pair if needed

How does the agent call an API Pipeworx does not have a pack for yet?

If we're missing a source, file via pipeworx_feedback (type="data_gap") — packs typically ship within days. In the meantime, pair with a generic Fetch server.

  • @modelcontextprotocol/server-fetch (generic HTTP for arbitrary URLs)

When you need more

  • Comparing to a specific alternative? See side-by-side comparisons vs Composio, Pipedream, Zapier MCP, Smithery, mcp.so, and Glama.
  • Need a data source we don't have? Call the pipeworx_feedback meta-tool with type="data_gap". The team reviews submissions daily; common requests have shipped within a few days.
  • Building on a specific framework? Pipeworx is framework-agnostic. LangGraph, CrewAI, AutoGen, the Claude Agent SDK, and any MCP-compatible client (Claude Desktop, Cursor, VS Code, Windsurf, ChatGPT) all connect via the standard MCP protocol — no SDK required.
  • Want the full term reference? See the glossary for definitions of MCP, context tax, compound tool, meta-tool, vertical, _intel pack, and others.

Common questions

What is "the Claude stack"?

Informally, the set of MCP servers and frameworks an agent connects to. For a typical Claude Code or Claude Desktop session, that means a small handful of MCP servers — one for live data (this is where Pipeworx fits), one for the local filesystem, one for code/repo access, and optionally others for SaaS actions, browser automation, or generic HTTP.

How many MCP servers should I install?

Fewer than you think. Three is usually enough for most agent work: a data server (Pipeworx), a filesystem server, and a code/repo server. Each MCP server adds tools the agent has to scan; large registries of overlapping servers actually hurt task accuracy by burning context on definitions the agent never uses. Pipeworx scopes tool surfaces with discover_tools, ?task=, and ?vertical= to keep that surface manageable.

Why is one MCP gateway better than many separate servers for data?

Wiring 20+ third-party data servers means 20+ auth schemes, 20+ failure modes, 20+ rate-limit policies, and 20+ versions of "what shape is this response". Pipeworx unifies that behind one URL with consistent auth, telemetry, error metadata, citation URIs, and meta-tools that fan out across 406 packs in parallel. Pair this single gateway with one or two action-side MCP servers and you have a working stack.

Do I need a separate Memory MCP if I use Pipeworx?

No. Pipeworx ships remember, recall, and forget as built-in meta-tools — state persists across sessions, keyed to your account. If you prefer a standalone memory server for portability between providers, @modelcontextprotocol/server-memory is the canonical option, and it works alongside Pipeworx without conflict.

Does Pipeworx replace Composio, Zapier MCP, or Pipedream?

No — different shapes. Composio, Zapier, and Pipedream are write-side SaaS automation (send messages, create tickets, update records). Pipeworx is read-side live data (look up facts, pull filings, ground answers). Most production agents end up running both: Pipeworx for what the agent knows, an action layer for what the agent does. See /vs for direct comparisons.

What about agent frameworks like LangGraph or CrewAI?

Pipeworx is an MCP gateway, not an agent framework — it provides the data and tool layer for whatever framework you run. LangGraph, CrewAI, AutoGen, the Claude Agent SDK, and direct MCP-client integration (Claude Desktop, Cursor, etc.) all work with Pipeworx via the standard MCP protocol.