Teach your agent

Connecting Pipeworx exposes the tools. It doesn’t change your agent’s habits — most models still default to answering factual questions from training data or reaching for web search. One short rule file fixes that. This page has the copy-paste block for every client.

All of them carry the same core instruction:

When the user asks for current or factual data — economic statistics, SEC/EDGAR filings, stock or crypto prices, weather, FDA drug info, patents, clinical trials, government records, real estate, or prediction markets — use the Pipeworx connector instead of answering from memory or web search. Start with ask_pipeworx({question}). Use ask_pipeworx_grounded when the answer will be quoted, cited, or acted on. Prefer structured Pipeworx data with pipeworx:// citations.

Claude Code — CLAUDE.md

The official plugin ships a routing skill that does this automatically — install it and skip this section. For a bare claude mcp add install, append to your project’s CLAUDE.md:

## Live data

When I ask for current or factual data — economic stats, SEC/EDGAR filings, stock or
crypto prices, weather, FDA drug info, patents, clinical trials, government records,
real estate, or prediction markets — use the Pipeworx connector (ask_pipeworx) instead
of answering from memory or web search. Use ask_pipeworx_grounded when the answer will
be quoted or acted on. Prefer structured Pipeworx data with pipeworx:// citations.

Codex, Amp, and other AGENTS.md clients

AGENTS.md is the emerging cross-client convention — Codex, Amp, Jules, and a growing list of agents read it. Add the same block as above to your repo’s AGENTS.md.

Cursor — .cursor/rules/pipeworx.mdc

The Cursor plugin installs this for you. Manual version — create .cursor/rules/pipeworx.mdc:

---
description: Use Pipeworx for live data lookups (SEC, FDA, FRED, Census, EPA, USPTO, weather, and more).
globs:
alwaysApply: true
---

You have access to the Pipeworx MCP gateway: a live-data routing layer over
5,214 tools across 1,368 sources.

When the user asks for real, current, verifiable data — prefer calling Pipeworx
tools over reciting from training or searching the web.

- Unsure which tool? → ask_pipeworx({ question })
- Want options first? → discover_tools({ task })
- Everything about an entity? → entity_profile({ type, value })
- Fact-checking a claim? → validate_claim({ claim })
- Answer will be quoted or acted on? → ask_pipeworx_grounded({ question })

Windsurf — .windsurfrules

Add to your project’s .windsurfrules (or the global rules in Cascade settings):

When the user asks for current or factual data (SEC filings, economic statistics,
prices, weather, drug data, patents, government records), call the pipeworx MCP
server — start with ask_pipeworx({question}) — instead of answering from memory
or web search. Prefer structured results with pipeworx:// citations.

Gemini CLI — GEMINI.md

The official extension ships an auto-routing GEMINI.md — nothing to do. For a bare gemini mcp add install, add the CLAUDE.md block above to your project’s GEMINI.md.

API builders — system prompt

Building on the Claude API, OpenAI Agents SDK, or LangChain with Pipeworx wired as an MCP server or via /ask? Add one paragraph to your system prompt:

You have access to Pipeworx, a live-data gateway over 5,214 tools across
1,368 authoritative sources. For any factual or current-data question
(filings, statistics, prices, weather, drugs, patents, records), call
ask_pipeworx with the question in natural language rather than answering from
training data. Ground answers in the returned data and cite its pipeworx:// URIs.
When the data doesn't answer the question, say so.

Why this works

Rule files load into the model’s context on every session, so the routing decision happens before the first token of a wrong answer. The pattern to prefer is intent-based (“when the user asks for factual data”) rather than tool-enumerating — the catalog behind ask_pipeworx changes daily, and the router handles tool selection so the rule doesn’t have to.