AI agent frameworks: 13 MCP clients, 5 servers, 6 A2A

13 clients. 5 servers. 6 A2A pages that match the package you can install. That is the split on November 13, 2026, across the 13 frameworks whose installable versions are in the table.

An MCP client lets the agent call a tool server. An MCP server lets some other client call the agent. A2A is a third wire, agent to agent, and it is the short list. AG-UI is a fourth, and most of the "support" is an adapter named on docs.ag-ui.com/integrations, not a module inside the framework package.

The November 5 GitHub snapshot is stars and one latest tag. Best AI agent frameworks is the licence and that tag. LangGraph vs CrewAI vs the OpenAI Agents SDK is three bills. None of those pages is this matrix.

Which installable version was current on November 13?

The registry version is what pip or npm would have resolved. The GitHub tag is what the releases API called latest. They match on LangGraph, CrewAI, Google ADK, Pydantic AI, Agno, AutoGen, LlamaIndex, smolagents, and the Vercel AI SDK package ai. They do not match on Microsoft Agent Framework, Mastra, or Strands.

Install the registry version when you are reproducing a bug. Quote the GitHub tag when you are talking about what GitHub called latest. Microsoft's 1.26.0 is a real upload. It is not, on this read, a GitHub release.

Which cells are actually documented?

Thirteen rows document an MCP client. Five document a server path. Six document A2A on the installable release. Yes means the framework's own page describes it. Listed means the AG-UI integrations page names it. A dash means the page in the notes did not document that feature. A dash is not a no.

LangGraph's client is MCPAdapter on the LangChain docs, which is the layer a LangGraph agent uses to call MCP servers. The server is a different product. Agent Server exposes LangGraph agents as MCP tools at /mcp, and the page says the protocol behavior it describes requires langgraph-api>=0.18.0. Streaming is stream_mode="messages". Human-in-the-loop is interrupt(). Tracing is LangSmith's integration, which is not the open-source graph package.

CrewAI 1.15.41's MCP page is servers as tools: the mcps field and MCPServerAdapter. That is a client. It does not document CrewAI itself as an MCP server. A2A is a different page in the same 1.15.41 docs: agents delegate to remote A2A agents and can run as A2A servers. Streaming, approval gates, and evals were not on either page, so those cells stay empty. AG-UI lists CrewAI.

The OpenAI Agents SDK MCP guide is also a client: hosted MCP, Streamable HTTP, SSE, and stdio. The streaming and human-in-the-loop guides are in the same docs. This read did not quote the tracing guide, so that cell stays empty. The Python package on PyPI is still 0.24.0 from October 15. The JavaScript package @openai/agents reached 0.22.0 on November 12.

Google ADK 2.14.1 uses McpToolset as the client, and the same MCP page says an ADK server can be configured as an MCP server for other clients. The A2A guide covers exposing an agent and calling a remote one. The confirmation page is the human-in-the-loop cell. A standalone streaming guide was not in the pages that resolved, so streaming stays empty even though the A2A material elsewhere talks about a streaming mode.

Microsoft Agent Framework's tools page has rows for MCP servers the provider runtime calls, and for local MCP tools. That is a client. The A2A serve page is the server direction and links a page for invoking a remote agent. AG-UI is documented by Microsoft, including pip install agent-framework-ag-ui and agent.run(..., stream=True). Workflow human-in-the-loop uses a RequestPort. The observability URL tried for this note returned 404, so tracing stays empty. The installable Python package is 1.26.0. GitHub's latest release tag is still python-1.25.0.

AutoGen's MCP reference wraps an MCP tool over STDIO. That is a client, on a package uploaded September 30, 2025. The AG-UI list names AG2, not AutoGen. This note does not treat those as the same row. The maintenance argument is Is AutoGen deprecated?. This matrix does not add a new claim about the README.

Pydantic AI 2.68.0 connects to MCP servers and can sit inside an MCP server. Deferred tools are the approval pause. A guessed A2A URL redirected away from Pydantic's docs, so A2A stays empty. AG-UI lists Pydantic AI.

Mastra's connections page documents both MCPClient and MCPServer, on MCP revision 2026-07-28. The built-in server is /api/mcp for agents, workflows, and tools already registered on the Mastra instance. The A2A page says Mastra supports protocol version 0.3.0, to expose agents as remote agents and to consume remote A2A agents as subagents. npm is 1.83.0. The latest GitHub release is 1.82.0 from November 4. AG-UI lists Mastra.

Agno 3.2.0's MCP overview connects an agent to an MCP server with MCPTools. That is a client. The overview does not document Agno as the MCP server. A2A is the AgentOS interface: the page says to expose Agno agents via the A2A protocol, and the version it names, v2.1.2, is the protocol. AG-UI lists Agno.

smolagents 1.26.0, unchanged since May 29, loads MCP tools with MCPClient and ToolCollection.from_mcp on the tools tutorial. It is not on the AG-UI list. The tutorial does not document smolagents as an MCP server. The versioned A2A URL says the page does not exist in v1.26.0 and exists on main. The package you install does not have that page.

Strands 1.64.0 loads MCP tools in Python and TypeScript, and its A2A page documents both consuming and serving. AG-UI lists Strands as AWS Strands Agents. The PyPI package and the repository's latest GitHub tag are different objects: 1.64.0 versus harness-typescript/v0.4.0.

The Vercel AI SDK at ai 7.0.171 connects to MCP servers. streamText is the streaming cell. The MCP page is the client direction. The AG-UI integrations list did not name the Vercel AI SDK. GitHub's release ai@7.0.171 matches npm, the same hour. The repository publishes other packages from the same tag list, so "latest release" on that repo is not always the ai package.

LlamaIndex 0.14.25's MCP example covers using tools from an MCP server and converting a workflow to an MCP app. AG-UI lists LlamaIndex. The example page does not document A2A. The package date is September 21, 2026, the oldest current release in this set after AutoGen and smolagents.

Where does AG-UI sit, if it is not the framework?

On the adapter shelf, unless the vendor's own docs install it.

The integrations page read November 13 names LangChain, not a heading that says LangGraph. LangChain is the MCP client layer this note already used for LangGraph. The protocol repository also ships a Python package, ag-ui-langgraph, which is an adapter you install beside LangGraph. That package is not langgraph 1.2.21.

Microsoft is the exception in this set. The Learn page tells you to install agent-framework-ag-ui and maps framework runs onto AG-UI events, including approval. That is first-party documentation of an integration package. It is still a separate install from agent-framework.

smolagents, AutoGen, and the Vercel AI SDK were not on the list. If a demo wraps one of them, this note did not find it on that page.

What does this mean for the framework you install?

Call MCP when the agent needs someone else's tools. Every framework in the table documents that, and the docs are not interchangeable. CrewAI wants an mcps list. ADK wants McpToolset. The OpenAI SDK wants a hosted tool or a transport class. Copying one project's snippet into another project is how the tool call fails before the model sees it.

Expose MCP when some other host, a coding agent or a second product, has to call your agent. The list is Agent Server, ADK, Pydantic AI, Mastra, and the LlamaIndex MCP app. LangGraph's open-source graph does not grow /mcp because the Agent Server page exists. You install the server.

Use A2A when the other party is an agent with its own card, not a tool with a JSON schema. On the installable docs, that is CrewAI, ADK, Microsoft Agent Framework, Mastra, Agno, and Strands. Agno's page is the expose direction. The other five describe both calling and serving. Do not describe the other seven as incompatible. LangGraph's /a2a URL returned 404. The OpenAI Agents SDK, LlamaIndex, and Vercel AI SDK URLs tried the same day returned 404. Pydantic's guessed URL left the docs for github.com/datalayer/fasta2a. AutoGen's guessed URL returned 404. smolagents is the one that says the page is missing from the release you can install.

Do not buy AG-UI support off a logo strip. Read the integrations page, then read the framework's own install line. Microsoft has one. The others on the list are the protocol project's adapters until their own docs say otherwise.

If the bill you actually need is the search call those tools make, that is a different page: web search API cost per agent task. Framework release tags from the November 5 pull are the GitHub stats report. How often those tags landed is the 30-day cadence note.

What should you open next?

The comparison shelf, the trackers, and the directory.

• Best AI agent frameworks is the licence and release-tag comparison. This note does not replace it.
• AI agent framework GitHub stats, November 2026 is the November 5 snapshot these names come from.
• Which AI agent framework do people install? is file requests, not protocol support.
• The downloads tracker and the stars tracker move after this page. A download is not a user.
• Web search API cost per agent task prices the search call once a framework can make one. All free tools are on one page.
• Submit or claim a listing if you ship an agent and want it in the directory.