Glossary

An open protocol for connecting AI agents to external tools and data. MCP standardizes how an agent discovers, calls, and cites data sources over JSON-RPC. The spec defines three primitives — Tools, Resources, and Prompts — that any MCP-compatible client (Claude Desktop, Claude Code, Cursor, VS Code, ChatGPT, etc.) can consume.

→ Full MCP, not just tool calling

An MCP primitive representing a callable function the agent can invoke — for example, fred_get_series or edgar_company_filings. Each tool has an inputSchema (often with examples), an outputSchema, and a description. Pipeworx exposes 5,462 tools across 1419 packs.

An MCP primitive for citable URIs that point to specific entities — e.g. pipeworx://edgar/company/0000320193/filings or pipeworx://clinicaltrials/study/NCT04568358. Resources let an agent embed durable references in its output instead of pasting raw data, and resources/read fetches current content for any URI.

→ Resources docs

An MCP primitive representing a pre-built workflow the agent can invoke by name — e.g. research_public_company(company) or drug_safety_profile(drug_name). Prompts encode optimal multi-step plans for common tasks so the agent does not have to rediscover the chain on every session.

→ Prompts docs

The single MCP server at https://gateway.pipeworx.io/mcp that fronts every pack. Agents connect to the gateway once and access the full catalog — no per-pack configuration. The gateway handles auth, rate limiting, routing, caching, telemetry, and meta-tools.

A bundle of tools for a single data source — for example, the edgar pack exposes SEC EDGAR filings, and the fred pack exposes FRED economic series. Packs are stateless and live in mcps/{slug}/src/index.ts in the repo. Pipeworx ships 1419 packs.

A curated bundle of packs optimized for a domain — Housing Intel, FinTech, PharmaPulse, GovCon Radar, AgriPulse, TradeGuard, and GreenLens. Connect to a vertical with ?vertical={slug} on the gateway URL to load only the relevant subset of tools rather than the full catalog.

→ Verticals docs

A pack-level tool that chains multiple underlying API calls into one response — for example, housing_property_report combines property detail, AVM valuation, sales history, and tax assessment in a single call. Compound tools collapse 5–15 round-trips into one.

→ Compound tools docs

A gateway-native tool that operates across the entire catalog rather than a single pack. Pipeworx ships seven: ask_pipeworx, discover_tools, resolve_entity, compare_entities, entity_profile, recent_changes, and remember/recall/forget.

→ Meta-tools docs

A naming convention for compound packs that fan out across multiple data sources for a specific domain — fintech-intel, trade-intel, pharma-intel, govcon-intel, housing-intel. Each _intel pack offers ~6 high-level workflows that combine its underlying sources.

Bring Your Own key — for premium sources like ATTOM, Altos Research, or Alpha Vantage, you pass your own API key via the X-API-Key header or ?apikey= query parameter. Public sources (FRED, SEC EDGAR, FDA, Census, BLS, EPA, ClinicalTrials, USPTO, etc.) require no keys at all.

The token cost of loading tool definitions into a model's context window for tools the agent never ends up calling. Loading 1,000 tool definitions burns model bandwidth on unused metadata. Pipeworx avoids the context tax by returning only the ~20 most relevant tools per task via discover_tools or vertical scoping.

→ Context tax docs

Metadata block attached to every Pipeworx tool response, containing cost (credits charged), cache (hit/miss, fresh_until timestamp), suggestions (what to call next based on session history), and alternatives (what to try if a call fails). Agents reason about _meta to decide whether to refetch or self-correct.

A stable resource URI scheme for citing entities returned by Pipeworx tools — e.g., pipeworx://edgar/company/0000320193/filings or pipeworx://fred/series/UNRATE. Resolvable via the MCP resources/read method, durable across sessions, and safe to embed in agent output as a citation.

Routes a natural-language question to the right tool, fills the arguments, executes it, and returns the answer. Best entry point when you do not know which pack or tool to call. Free for all tiers.

Returns the top-20 most relevant tools for a given task description, with full schemas. Replaces the pattern of dumping all tool definitions into context.

Given a type ("company" or "drug") and value (ticker, name, etc.), returns canonical IDs across packs in one call — CIK + ticker + LEI for companies, RxCUI + brand + ingredient for drugs. Returns ready-to-embed pipeworx:// resource URIs.

Side-by-side comparison of 2–5 entities in parallel. For companies, returns revenue + income + cash + debt with periods. For drugs, returns interactions, indications, and adverse-event summaries.

Fan-out fetch for "everything about this entity" — for a company, returns SEC filings + fundamentals + patents + federal contracts + recent news + LEI in a single round trip across 6 packs.

Time-filtered fan-out — "what is new about this entity in the last N days/months". Accepts since as ISO date or relative shorthand (7d / 30d / 3m / 1y). Returns structured items with total_changes count. Designed for monitoring workflows.

Agent memory across MCP sessions. remember stores a note keyed by topic; recall retrieves notes by topic or search; forget deletes them. State persists on the gateway, keyed to the agent's account or session ID.

→ Memory docs

No authentication required. Rate-limited to 50 calls per day per IP address (CF-Connecting-IP). Sufficient for evaluation and casual use. Public data sources only.

Authenticated via X-API-Key header or ?apikey= query parameter — the key is for an upstream paid source (ATTOM, Alpha Vantage, etc.) rather than a Pipeworx account. Rate-limited to 200 tool calls per day per IP + key hash.

Free Pipeworx account linked via GitHub OAuth. Bearer token authentication. Rate-limited to 200 calls per day per account. Required for verticals.

Stripe-billed Pipeworx subscription. Bearer token authentication. Unlimited calls, metered per call. Includes access to all premium verticals (Housing Intel, FinTech, PharmaPulse, etc.) and team-level features.

The single billing unit for everything on Pipeworx, pegged at 1 credit = $0.0001. Requests, answers and metered vendor calls all settle in credits, so there is one currency to reason about rather than a separate price per product.

→ Pricing

One billable request against your calendar-month volume counter — the unit the volume brackets are measured in. It is "anchored" because the price of a given call depends on where it falls in the month: the 9,000th request of a month is charged at a different bracket rate than the 30,000th.

Requests are priced by monthly volume, marginal like tax brackets rather than a cliff: 50 credits each for the first 10,000 anchored requests in a calendar month, 30 to 20,000, 20 to 50,000, then 10 beyond. Crossing a threshold reprices only the requests above it, never the ones already made.

→ Pricing

An ask_pipeworx-class call is not billed at a flat rate — it settles at the measured inference cost multiplied by 1.5, plus one anchored request for the underlying tool call. You pay for the reasoning actually spent on your question rather than an average.

One of the few packs fronting a vendor that charges per unit. These settle at the vendor's real unit cost times a volume markup — 1.5x under 10,000 calls a month, then 1.4x, 1.3x, and 1.2x past 50,000 — floored at 100 credits. Bring your own vendor key and Pipeworx adds nothing on top.

A per-account cap on how many credits a single day can consume, checked before the balance is drawn down. It bounds the damage from a runaway agent loop, which a monthly budget alone cannot.

→ Account