Prompts

The MCP spec’s third noun is Prompts: pre-baked workflow templates the server delivers to the agent, with arguments substituted in. The agent runs the prompt in its own context, executing the tool sequence the server suggests.

Most MCP servers ship Tools and stop. Pipeworx ships 12 Prompts — opinionated playbooks for the highest-value cross-pack workflows.

Why server-side prompts

Agents are decent at planning multi-step tool calls but they rebuild the plan from scratch every session. The same agent solves “research a public company” with slightly different tool sequences each time, missing one or two relevant calls per run.

Prompts encode the optimal sequence once, on the server. Every agent that calls prompts/get gets the same proven plan with their arguments filled in.

How an agent uses a prompt

// Discover available prompts
prompts/list()
// → [{ name: "research_public_company", description: "...", arguments: [{name: "company", required: true}] }, ...]

// Get a prompt with arguments substituted
prompts/get({
  name: "research_public_company",
  arguments: { company: "AAPL" }
})
// → { messages: [{ role: "user", content: { type: "text", text: "Research the public company AAPL using Pipeworx. Use this sequence:\n\n1. Resolve the company...\n2. Get recent filings...\n3. Pull key financial concepts...\n..." }}] }

The returned message is what the agent reads as its task. It contains the suggested tool sequence with the company filled in. The agent then executes it.

Available prompts

Prompt	Args	Tools orchestrated
`research_public_company`	`company`	EDGAR + CFPB + FRED
`drug_safety_profile`	`drug_name`	RxNorm + FDA + ClinicalTrials
`trade_exposure`	`reporter`, `partner`	UN Comtrade + Census + Treasury
`housing_market_brief`	`metro`	FRED + ATTOM + Altos + BLS
`regulatory_landscape`	`company`	SEC + Federal Register + EPA + USAspending
`property_research`	`address`	ATTOM + Altos + FEMA
`patent_due_diligence`	`subject`	USPTO + EDGAR + Trademarks
`lobbying_activity`	`company`	Lobbying + FEC + USAspending + Federal Register
`financial_health_check`	`ticker`	resolve_entity + EDGAR + CFPB + insider trading
`agricultural_commodity_brief`	`commodity`	USDA NASS + BLS + FRED + Comtrade + weather
`environmental_risk`	`subject`	EPA ECHO + GHG + TRI + AirNow + Federal Register
`tool_chain_for_question`	`question`	meta-prompt — uses discover_tools to plan

When to use a prompt vs. ask_pipeworx

ask_pipeworx is one round trip: question in, answer out. Best for “what is X” questions where one tool gives the answer.
A prompt is a multi-step plan: the server hands the agent a sequence to execute. Best for “research X end-to-end” tasks that span 3–5 tools.

If the task fits one tool, use ask_pipeworx. If it needs synthesis across packs, get a prompt.

Customizing the playbook

Prompts are static at the gateway level — every agent gets the same template. If you need a custom workflow:

Use discover_tools to find the relevant tools
Compose your own prompt or chain the calls directly
File feedback via pipeworx_feedback({type: "feature", message: "Add prompt for X"}) — the team adds prompts based on signal

The bar to add a new prompt is low. If you’re rerunning the same 3-tool sequence repeatedly, that’s a candidate.