Niches · Agent frameworks & SDKs

Give an agent tools from MCP servers

One company's docs show a full fit.

1companies, full fit
17companies, fit with a limit
93companies carry the tag
12share of the category's work
9first-hand accounts

What counts as done

The agent lists and calls tools from two or more MCP servers at once, with name clashes handled, auth headers set per run or per user, tool errors passed to the model instead of crashing the run, and servers that fail to connect or need approval reported to the host.

Measured, for our own tests: connect time for 3 servers and run success with concurrency above one. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: connect the GitHub MCP server to my agent, use these MCP servers from my Python agent, add per-user auth to the MCP calls, load tools from MCP as needed

Does the job 1

The docs show it meets the test above, and nothing we read says otherwise.

ListingWhy
Claude Agent SDK BB
Anthropic, PBC
stdio, HTTP, SSE and in-process servers at once, tools namespaced mcp__server__tool, headers set per query, and the init message reports each server as connected, pending, failed or needs-auth. Measured connect time is the suite's. source

Does the job, with a limit 17

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
Agno BB
Agno Inc.
MCPTools and MultiMCPTools take stdio, streamable HTTP and SSE, tool_name_prefix avoids clashes, and header_provider sets headers before each run. How a server that fails to connect is reported to the host isn't shown on the pages read. source
BoxLang AI B
Ortus Solutions, Corp.
mcpServers takes HTTP and stdio servers with a prefix against clashes, headers, bearer and OAuth. A server that fails to start is only written to the ai log and the agent carries on, so the host isn't told; the stdio allow-list is open by default. source
CrewAI C
CrewAI, Inc.
MCPServerAdapter aggregates stdio, SSE and streamable HTTP servers at once, but the page shows no tool-name prefix for clashes or per-run headers, and a failed server raises from the context manager. source
Cursor Agent SDK B
Anysphere, Inc.
Connections reach HTTP and stdio servers, with credential templates in headers and a per-session auth callback for HTTP. An account found servers needing approval vanish from the SDK with no event (forum 178033). source
Docker Agent BB
Docker, Inc.
The mcp toolset takes local stdio, Docker MCP catalog and remote streamable HTTP or SSE servers, with OAuth handled automatically. The page shows no tool-name prefix for clashes or per-run headers for many users. source
Agent Development Kit (ADK) BB
Google LLC
McpToolset takes stdio and HTTP connections with tool_filter. The MCP page read shows no name-prefix option, per-user headers or failure reporting, and the listing notes no error handling on the page. source
guilds.run C
Cigale AI
Connectors are remote MCP endpoints signed in per organisation, guild or agent with token or OAuth, plus custom headers. Only remote HTTPS endpoints are described, configured in the app rather than per run in code. source
LangGraph BB
LangChain, Inc.
LangChain's MCP adapter connects several servers, prefixes each tool with its config key and mixes stdio and HTTP; the listing records langchain.mcp as beta. Failure reporting to the host isn't described on the connections page. source
LlamaIndex Framework BB
LlamaIndex, Inc.
MCP tools load from existing servers into workflows and workflows can be served as MCP. The guide read gives no name-clash handling or failure reporting, and an account found no hook to set tool inputs from the authenticated user (#20386). source
Mastra A
Kepler Software, Inc.
MCPClient connects several servers; listTools() prefixes tools with the server name, but the per-user listToolsets() path uses raw names that can collide. Headers and OAuth are supported; how unreachable servers are reported to the host wasn't shown. source
Microsoft Agent Framework A
Microsoft Corporation
MCPStdioTool and MCPStreamableHTTPTool connect local and remote servers, headers can be set per run or through a header_provider, and approval_mode applies to MCP tools. Name-clash handling and failure reporting weren't shown on the page. source
NitroStack C
Nitrostack Inc.
Covers the reverse case: exposing the team's tools as an MCP server over STDIO and HTTP, with guards for JWTs and API keys. It is not an MCP client, so an agent can't use it to call other servers. source
OpenAI Agents SDK AA
OpenAI OpCo, LLC
Hosted, streamable HTTP, SSE and stdio servers; include_server_in_tool_names prefixes tools (opt-in); failed servers are dropped and tracked in failed_servers with reconnect(). Headers are set when a server is created, so per-user auth (#1969) still needs one connection per user. source
Pydantic AI AA
Pydantic Services Inc.
MCPToolset takes Streamable HTTP, SSE and stdio, .prefixed() avoids name clashes, tool errors go back to the model as ModelRetry by default, and headers or a custom http_client are set per toolset. An account hit a cancel-scope error with concurrency above one (pydantic-ai #2818). source
smolagents B
Hugging Face, Inc.
MCPClient loads tools from stdio or streamable HTTP servers, including several at once. The tools page shows no name-prefix option, per-run headers or failure reporting. source
Strands Agents A
Amazon Web Services, Inc.
MCPClient takes stdio and streamable HTTP (including OAuth or IAM), prefix_with_server_name for clashes, and tool_filters. Per-run headers and failure reporting weren't shown on the page. source
AI SDK AA
Vercel Inc.
createMCPClient supports HTTP (recommended), SSE and stdio (local only), with headers in the transport. No name-prefix option for clashes appeared on the page, and an open account says prepareStep can't change the tool list per step (vercel/ai #9743). source

Doesn't do this job 2

Listings in the category that do something else, with the reason.

ListingWhy
AgentOS A
Framers Lab, Inc.
No MCP client page in the docs sitemap or llms.txt; tools come from the app, extensions or the forge. source
TradingAgents B
Tauric Research
No MCP client or server; it is driven through Python or the CLI. source

What stops agents finishing

  • two servers exposing the same tool name stop the agent
  • headers fixed at start-up, so one server connection cannot serve many users
  • no hook to set tool arguments from the authenticated user before a call
  • long MCP tasks deleted by a TTL while still working
  • no MCP client at all in some listings

Around it

Before: agent-auth.connect-user-account. After: frameworks.tool-calling-agent.

Where this comes from

Weight 12 of 100 in Agent frameworks & SDKs: 7 of 61 accounts have this as their main job (share 11.5, rounded by largest remainder to sum 100)

The same as data: /api/v1/use-cases/frameworks.mcp-tools.json, or the MCP tool get_use_case with id frameworks.mcp-tools. Other jobs in this category: /api/v1/niches?category=frameworks.

Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.