AgentSidekick

AI agent product types

AgentSidekick sorts 73,820 indexed products into 16 formsDER. MCP servers are the largest at 69,078 (93.6%); 9,469 products across all types have full profiles. The forms differ in what you install and who runs it: a platform hosts, an MCP server is a running service, a skill is text the agent reads, a framework is code you build inside. As of 2026-11-21.

Every type, by size

Indexed = every active record. With a profile = records complete enough for their own page.

Product types with indexed and profiled counts, as of 2026-11-21
TypeHow it differsIndexed ShareWith a profile Researcher checked
MCP serversA running service the agent calls; a skill is text the agent reads.69,07893.6%5,5410
Agent skillsAdds instructions and procedure to the agent; an MCP server adds callable tools over a protocol.2,0562.8%1,8980
PluginsA container for skills, commands and MCP configuration, tied to one host format.1,0881.5%7230
InfrastructureRuntime plumbing underneath agents; frameworks orchestrate, infrastructure hosts and observes.3840.52%36441
FrameworksOwns the control flow; an SDK is a binding you call from your own code.2730.37%24826
Rules packsAlways loaded and passive; a skill is loaded on demand and can ship scripts.2670.36%1010
Agent projectsRuns a task itself; a framework is what such a project is often built with.2440.33%1950
IDE extensionsPackaged for an editor's extension system; a plugin targets an agent host.1480.20%1280
LibrariesOne component; a framework supplies the whole loop.1160.16%10526
PlatformsHosts extensions rather than being one. Compatibility is recorded per platform.470.06%4747
AI APIsA remote endpoint without the MCP layer; an MCP server often wraps one.290.04%2929
SaaS productsBought and used as a finished product; an API or framework is a building block.260.04%2626
Developer toolsUsed during development; infrastructure serves the running agent.230.03%2323
Agent SDKsBinds one API or protocol; a framework provides a general agent architecture.150.02%1515
Protocols & standardsA specification, not software; SDKs and servers implement it.140.02%1414
Automation platformsConfigured, not coded; a framework is code.120.02%1212

Definitions

Editorial definitions of each form. Records are typed by the source registry where it says, otherwise inferred from repository topics.

Platforms 47 indexed · 47 profiled

A product people run agents in: an AI assistant, coding agent, IDE, command-line agent or agent builder. Platforms are hosts. Other products extend them through mechanisms the vendor documents, such as MCP servers, Agent Skills, plugins or rules files.

Versus its neighbours: Hosts extensions rather than being one. Compatibility is recorded per platform.

Most notable: Continue, OpenClaw, OpenCode

Agent projects 244 indexed · 195 profiled

A standalone agent project: a repository that is itself an agent doing a task end to end (research, coding, browsing, trading) and runs on its own. It is neither a host for extensions nor a toolkit for building agents.

Versus its neighbours: Runs a task itself; a framework is what such a project is often built with.

Most notable: graphify, orca, ECC

Plugins 1,088 indexed · 723 profiled

An installable bundle for one host platform that packages several extension types (slash commands, subagents, hooks, Agent Skills, MCP server configuration) behind a marketplace manifest. The format belongs to the host, so a plugin is not portable the way an MCP server or skill is.

Versus its neighbours: A container for skills, commands and MCP configuration, tied to one host format.

Most notable: mattpocock-skills, superpowers, hyperframes

Agent skills 2,056 indexed · 1,898 profiled

A folder with a SKILL.md file (front matter giving a name and description, then Markdown instructions) and optional scripts and reference files. An agent reads the description up front and loads the full skill only when a task matches. The format is the open Agent Skills standard.

Versus its neighbours: Adds instructions and procedure to the agent; an MCP server adds callable tools over a protocol.

Most notable: open-design, Understand-Anything, career-ops

MCP servers 69,078 indexed · 5,541 profiled

A program that exposes tools, resources and prompts over the Model Context Protocol (JSON-RPC over stdio or streamable HTTP), so any MCP client can call it. It runs locally as a process or remotely as a hosted endpoint, and usually wraps one service or data source.

Versus its neighbours: A running service the agent calls; a skill is text the agent reads.

Most notable: REA, World Monitor, HyperFrames by HeyGen

Frameworks 273 indexed · 248 profiled

A code library that supplies the agent loop itself: orchestration, state, memory, tool calling and multi-agent coordination. You write your application inside its structure, and it is usually model-agnostic.

Versus its neighbours: Owns the control flow; an SDK is a binding you call from your own code.

Most notable: LangChain, RAGFlow, AutoGen

Agent SDKs 15 indexed · 15 profiled

A software development kit maintained for one vendor API or one protocol, with typed client and server bindings (for example the official MCP SDKs, or a model vendor's agents SDK). Narrower than a framework and tied to the API it binds.

Versus its neighbours: Binds one API or protocol; a framework provides a general agent architecture.

Most notable: OpenAI Agents SDK, AI SDK, Stagehand

Protocols & standards 14 indexed · 14 profiled

A published specification for how agents, tools, clients or services exchange messages or files: wire formats, message types and expected behaviour. Implementations (SDKs, servers, clients) are separate products, and one protocol can have many.

Versus its neighbours: A specification, not software; SDKs and servers implement it.

Most notable: Agent Skills, Agent2Agent Protocol (A2A), AGENTS.md

Rules packs 267 indexed · 101 profiled

A packaged set of instruction files (AGENTS.md, CLAUDE.md or editor rule files) that tell a coding agent the conventions of a codebase, language or stack. Plain text read into context on every run; nothing executes.

Versus its neighbours: Always loaded and passive; a skill is loaded on demand and can ship scripts.

Most notable: awesome-cursorrules, agency-agents-zh, ai-setup

Infrastructure 384 indexed · 364 profiled

Services and components agents run on without being agents: sandboxes and code execution, browser automation services, memory and vector stores, model gateways, observability, tracing and evaluation.

Versus its neighbours: Runtime plumbing underneath agents; frameworks orchestrate, infrastructure hosts and observes.

Most notable: Daytona, LiteLLM, Milvus

AI APIs 29 indexed · 29 profiled

A hosted HTTP API that agents call for one capability, such as search, scraping, inference or payments, keyed and usually billed by the provider. You get an endpoint, not code to run.

Versus its neighbours: A remote endpoint without the MCP layer; an MCP server often wraps one.

Most notable: Firecrawl, OpenAI API, Supermemory

Libraries 116 indexed · 105 profiled

A reusable code package for one part of agent work (document parsing, retrieval, tool wrappers, structured output, prompt handling) that does not supply an agent loop.

Versus its neighbours: One component; a framework supplies the whole loop.

Most notable: MarkItDown, Browser Use, MCP Reference Servers

SaaS products 26 indexed · 26 profiled

A hosted commercial product with its own interface, used by people rather than called by developers, where agent functionality is the product (for example an AI support or sales agent).

Versus its neighbours: Bought and used as a finished product; an API or framework is a building block.

Most notable: Composio, Polar, Pydantic Logfire

Developer tools 23 indexed · 23 profiled

Tooling for the people building agents: command-line utilities, MCP inspectors, testing and debugging tools, registries and package managers. Used while building, not at runtime.

Versus its neighbours: Used during development; infrastructure serves the running agent.

Most notable: Context7, Chrome DevTools MCP, Playwright MCP

Automation platforms 12 indexed · 12 profiled

A workflow product, often visual and node-based, where users chain triggers, app connectors and model calls into automations, with agent steps as one kind of node.

Versus its neighbours: Configured, not coded; a framework is code.

Most notable: n8n, Dify, Langflow

Limits

Each record has exactly one type. Registry records (the official MCP Registry, Open VSX, plugin marketplace manifests, SKILL.md files) take the type the registry implies. Auto-discovered GitHub repositories are typed from their topics, which is a heuristic and is wrong for some records; researcher-checked records (259 in all) are typed by hand.

Counts measure what public sources expose, not market size. MCP servers come largely from the official MCP Registry; skills, plugins, frameworks and the rest are found through GitHub, marketplace manifests and curated research, so the forms are not discovered with equal completeness. See the methodology for each source.

Page data last updated .