Meta MCP Glossary: Definitions for Every Term in the MCP Space

MCP is the Model Context Protocol — an open specification introduced by Anthropic in November 2024 that lets AI clients connect to external tools, data, and prompts through a uniform JSON-RPC interface. The full disambiguation of what people mean by Meta MCP sits in the Pillar; this glossary defines every term that shows up there and across the rest of the guide.

MCP foundation terms

Model Context Protocol (MCP). An open specification published by Anthropic at modelcontextprotocol.io^↗ in November 2024. It defines a JSON-RPC interface for AI clients to discover and invoke tools, fetch resources, and inject prompts — running over stdio, HTTP, or Server-Sent Events transports. The spec is Apache-2.0 licensed and vendor-neutral, which is why an MCP server built for Claude works inside ChatGPT, Cursor, or Continue without modification.

MCP server. A process that speaks MCP and exposes tools, resources, or prompts to AI clients. The server is the side that implements what AI can do — read a calendar, query a database, manage Meta ads campaigns. A single AI client can connect to many MCP servers in parallel; each contributes its own tools to the merged tool catalog the AI sees.

MCP client. The AI-facing side of the connection. The client speaks MCP back to the server, discovers available tools, and invokes them on behalf of the AI model. Claude Desktop, Cursor, Continue, and ChatGPT (with MCP support enabled) are all MCP clients. The client is not the AI model itself — it is the application embedding the model that handles the protocol.

MCP host. The application running the AI model and the MCP client together. In casual usage "client" and "host" are often conflated; the spec keeps them separate because a host can run multiple clients (one per server connection), each with its own permission scope.

Transport. How MCP messages travel between client and server. Three transports are defined: stdio (server is a local process, client pipes JSON-RPC over its standard input/output — fastest, most common for desktop installs), HTTP (request/response, server is a remote endpoint), and SSE / streamable HTTP (server pushes streamed responses, used for tool calls that produce long output). A given server can support more than one transport.

Tool. A function the AI can call through the server, with a defined input schema, output schema, and natural-language description. For Meta ads MCPs, tools include get_campaigns, create_ad_set, search_meta_interests, and similar — each one is a discrete operation the AI can invoke.

Resource. Read-only data the server exposes for the AI to fetch by URI — a file, a database row, a remote document. Distinct from tools, which can have side effects.

Prompt. A reusable prompt template the server publishes so the AI client can invoke it on demand. Used less often in ads MCPs; more common in dev-tool MCPs that bundle workflow templates.

Meta-specific terms

Meta Marketing API. Meta's official Graph API surface for managing advertising programmatically — campaigns, ad sets, ads, creatives, audiences, insights, catalogs, conversion pixels, and Page assets. Documented at developers.facebook.com/docs/marketing-apis^↗. Every Meta MCP server, including Meta's own, wraps the Marketing API underneath; the differences between MCPs are mostly about which slice of the API they expose, how they handle authentication, and what they bolt on top.

Meta ads MCP server. Meta's own MCP server, the Meta-hosted half of the "Meta ads AI connectors"^↗ launch on April 29, 2026. Free during open beta, no Developer App required, OAuth flows directly through Meta Business. Hosted by Meta, not open-source. Often called "Meta Official MCP" in search queries; that phrasing is community shorthand, not Meta's marketing label.

Meta ads CLI. The companion command-line binary launched alongside Meta's MCP server. Python 3.12+, distributed via pip or uv^↗, exposing the same Marketing API surface but as shell commands rather than MCP tools. The MCP server targets AI clients (Claude Desktop, ChatGPT); the CLI targets terminal sessions, scripts, and CI pipelines. They are not the same product despite frequent conflation in early write-ups.

Meta ads AI connectors. The umbrella label for Meta's April 2026 launch — the MCP server plus the CLI together. The launch announcement uses this term; "Meta MCP" is the search query users actually type.

Business Manager OAuth. Meta's authentication flow for ad accounts. The user logs in with their Meta credentials, selects which ad account and Page to authorize, and the MCP server receives a User Access Token plus, derived from it at runtime, Page Access Tokens for any Pages the user manages. All Meta MCPs that use official auth (Meta's own server, Pipeboard, MarketingRoutine) route through Business Manager OAuth.

Page Access Token. A short- or long-lived token derived from the User Access Token, scoped to a specific Page. Required for operations that act on Page assets — Instagram comment management, Page-owned creatives, organic posts.

Ad account. Meta's billing and asset container for advertising. Identified by an act_ prefix (e.g., act_1234567890). Each ad account belongs to a Business Manager and can be shared with multiple Pages and users with role-scoped permissions.

System user. A non-human Meta Business account used for server-to-server authentication, typically by agencies and large advertisers running automation that should outlive any individual employee's access.

Third-party Meta MCP terms

Disclosure before the named list: MarketingRoutine is the product the team publishing this guide works on. It appears below alongside competitors. The same definition format applies to every entry, and the Meta MCP Pillar goes deeper on the comparison surface.

Pipeboard. The most-starred open-source Meta Ads MCP — github.com/pipeboard-co/meta-ads-mcp^↗ — written in Python, exposing 29 tools across the Meta Marketing API. Licensed under Business Source License 1.1^↗, converting to Apache-2.0 on January 1, 2029. Available as a hosted endpoint at mcp.pipeboard.co/meta-ads-mcp or as a local install via pip or brew install pipeboard-co/tap/pipeboard.

MarketingRoutine. Commercial SaaS Meta Ads workflow layer, Meta-only, with a V2.1 beta focused on onboarding, workspace dashboard, MCP endpoint, bulk creative upload, and creative performance with thumbnails. Self-hosting not offered — the disclosed weakness against Pipeboard. Multi-platform not offered — the disclosed weakness against broader ad-suite products.

Madgicx. A commercial multi-platform suite that bundles Meta with other ad channels and layers MCP capabilities on top of an existing optimization platform. Closer to a managed-optimization service with MCP access than an MCP-native server. The right choice when multi-platform breadth is the binding constraint.

Ryze AI. Newer commercial entrant focused on agencies managing multiple Meta clients. Differentiating positioning: per-client token isolation, white-label client portals, creative-fatigue handling, and Business Manager auto-discovery. Less raw API breadth than Pipeboard or MarketingRoutine; more agency-shaped operational features.

brijr/meta-mcp. Open-source TypeScript implementation at github.com/brijr/meta-mcp^↗ — 25 tools, MIT-licensed, with npm install -g and one-click Vercel deployment options. A clean reference implementation for teams that want a TypeScript codebase to fork.

ZuckerBot. Listed in some MCP directories as a community-maintained Meta Ads MCP. The canonical repository URL has not been verified for this guide as of the lastVerified date; readers running ZuckerBot in production are invited to email a verified link so it lands in the next changelog entry.

Aggregator and meta-tool terms

MetaMCP. An MCP tool aggregator — software that fronts multiple MCP servers behind a single endpoint, so an AI client connects once and gets a merged tool catalog drawn from many backends. The "Meta" in MetaMCP refers to meta (above, across), not Meta Platforms. The naming collision is a recurring source of search confusion.

MCP proxy / OpenMCP gateway. Generalizations of the aggregator pattern. A proxy sits between client and server, often adding authentication, rate limiting, observability, or per-tool ACLs. A gateway typically refers to a multi-tenant proxy serving multiple downstream MCPs to multiple AI clients with policy enforcement.

Licensing and operational terms

Apache 2.0. Permissive open-source license used by the MCP specification itself and the reference SDKs. Allows commercial use, modification, and redistribution without copyleft requirements.

MIT. The most permissive widely-used open-source license. Used by brijr/meta-mcp. Allows everything Apache 2.0 does, with shorter terms.

BSL 1.1 (Business Source License 1.1). A "source-available" license — code is published, modification and use are allowed, but production use that competes with the licensor's commercial offering is restricted until the license converts (typically four years after release). Used by Pipeboard; converts to Apache 2.0 on January 1, 2029. MariaDB^↗ originated the license.

Source-available. Umbrella term for licenses (BSL, SSPL, Elastic License) that publish source code but restrict commercial use. Distinct from classic open-source under OSI-approved licenses. Worth flagging in compliance reviews because some buyers' policies allow only OSI-approved licenses.

Self-host. Running an MCP server on your own infrastructure, with full source access and no dependence on a vendor-hosted endpoint. The opposite of SaaS. Pipeboard and brijr/meta-mcp both support self-host; MarketingRoutine, Madgicx, Ryze, and Meta's own server do not.

SaaS. Software-as-a-Service — vendor-hosted, accessed by URL, paid as a subscription. MarketingRoutine, Madgicx, Ryze AI fit this shape. Meta's own MCP server is also vendor-hosted but free during beta.

Multi-tenant. A single deployment of the MCP server serving multiple isolated client accounts (typically agency clients), with token, data, and permission separation enforced at the application layer. Critical for agency operations; nice-to-have for single-brand setups.

Scheduled account monitoring. Server-side checks that run without an explicit AI prompt — creative-fatigue scoring, budget anomaly detection, performance regression flags. This is a common direction for ads tooling, but it is not part of MarketingRoutine's current V2.1 beta surface.

Rate-limit governance. Active management of the Meta Marketing API rate-limit headers (x-app-usage, x-business-use-case-usage) — throttling at high-usage thresholds, prioritizing user requests over background sync, retrying on 429. Important under heavy load; missing in most lightweight MCPs.

AI client terms

Claude Desktop. Anthropic's desktop application, MCP-native since the protocol launch. The most common host for Meta ads MCP servers in marketer setups.

ChatGPT MCP. OpenAI's MCP support, available in select ChatGPT tiers. Connects to remote MCP servers via HTTP or SSE transports.

Cursor. AI-first code editor with MCP support. Less common for marketer workflows; more common when MCPs are used for engineering tasks.

Continue. Open-source AI coding assistant, MCP-capable. Used in similar contexts to Cursor.

The FAQ below answers the most common standalone questions about MCP, the Model Context Protocol, and the Meta-specific terminology covered above.