By MV2 of Munim, Inc., with AI agents. Written on 12 November 2026 from the source of the MCP Python SDK (v2.3.0 and v1.27.2), the MCP TypeScript SDK (v2.4.0 and 1.32.2), the OpenAI Agents SDK (main branch that day), Pydantic AI v2.68.0 and Open WebUI v0.13.3. Nothing was run; see limits.
Your agent or chat app starts an MCP server and fails before any tool is listed:
mcp.shared.exceptions.MCPError: Request 'initialize' timed out
The client sent the MCP handshake and the server did not answer in time. Almost always the server is slow to start, not broken: npx or uvx downloading the package on first launch, a Python server importing heavy libraries, a server that logs in somewhere before it answers, or a busy CI runner. The fix is to give the handshake more time in your client's setting, and the setting, its name and its default differ a lot between clients. That is what this page lists.
Which text you see depends on the SDK version under your client, not on the server.
| Text you see | Where it comes from | Notes |
|---|---|---|
Request 'initialize' timed out (or another method in the quotes: 'tools/list', 'tools/call') | MCP Python SDK 2.x. Built as f"Request {method!r} timed out" in mcp/shared/jsonrpc_dispatcher.py (read at v2.3.0), raised as MCPError with code REQUEST_TIMEOUT (-32001). | The method name tells you which step was slow. |
Timed out while waiting for response to ClientRequest. Waited 5.0 seconds. | MCP Python SDK 1.x (read at v1.27.2, mcp/shared/session.py), raised as McpError with code 408. | No method name: ClientRequest is the wrapper class used for every client request, so the same text appears for the handshake, tools/list and tool calls. Look at whether any tool was listed before it. |
MCP error -32001: Request timed out | MCP TypeScript SDK 1.x (read at 1.32.2): McpError with ErrorCode.RequestTimeout (-32001); the class adds the MCP error <code>: prefix. | Used by many Node-based clients and IDEs. |
Request timed out (an SdkError with code RequestTimeout) | MCP TypeScript SDK 2.x (read at v2.4.0, packages/core-internal/src/shared/protocol.ts). | We did not check how your client prints SdkError. |
Failed to connect to MCP server '<id>' | Open WebUI. The timeout text above is in the backend log under it. | See that page for the other causes. |
| Client (version read) | Setting that bounds the handshake | Default | Also applies to |
|---|---|---|---|
MCP Python SDK 2.x, ClientSession or Client (v2.3.0) | read_timeout_seconds (float, seconds) | None: no timeout at all | Every later request that doesn't pass its own timeout, including call_tool (its read_timeout_seconds defaults to the session's). See also the 10-second probe. |
MCP Python SDK 1.x, ClientSession (v1.27.2) | read_timeout_seconds (a datetime.timedelta) | None: no timeout | Same: a per-request value wins, otherwise the session value is used. |
| MCP TypeScript SDK 1.x / 2.x (1.32.2, v2.4.0) | client.connect(transport, { timeout }) in milliseconds; in 1.32.2 connect() passes these options to the initialize request | DEFAULT_REQUEST_TIMEOUT_MSEC = 60,000 (60 s), per request | Each request takes its own timeout; resetTimeoutOnProgress (default false) and maxTotalTimeout (default none) exist for long tool calls. |
OpenAI Agents SDK, MCPServerStdio, MCPServerSse, MCPServerStreamableHttp (main, 12 Nov 2026) | client_session_timeout_seconds | 5 seconds in all three classes | It is passed to the SDK as the session read timeout (a timedelta with mcp 1.x, a float with mcp 2.x), so with the Python SDK rules above it is also the default for later requests. None or 0 turns it off. Per openai-agents-python#5405, the stock stdio example can fail with Request 'initialize' timed out on a cold start. |
Pydantic AI, MCPToolset (v2.68.0) | init_timeout: "Timeout in seconds for the initial connection and initialize handshake" | 5 seconds | Separate from read_timeout (default 300 s) for the long-lived connection. Both go to the FastMCP Client; if you pass your own pre-built FastMCP client, set them on that client instead (Pydantic AI raises an error if you pass either together with one). We did not read FastMCP's own timeout message. |
| Open WebUI (v0.13.3) | Environment variable MCP_INITIALIZE_TIMEOUT | 10 seconds | Passed as the SDK client's read_timeout_seconds, so it bounds the handshake and listing tools. Tool calls use AIOHTTP_CLIENT_TIMEOUT_TOOL_SERVER instead (no timeout when unset; utils/mcp/client.py). |
The MCP Python SDK itself has no handshake timeout. When a Python client gives up after 5 or 10 seconds, that number comes from the framework or app on top of it, so that is where to raise it.
npx -y some-server and uvx some-server may fetch the package on first run or after a cache clean, and that counts against the handshake. Installing it once (npm install -g, uv tool install) and pointing the config at the installed command takes that out.client_session_timeout_seconds=30 (OpenAI Agents SDK), init_timeout=30 (Pydantic AI), MCP_INITIALIZE_TIMEOUT=60 (Open WebUI, then restart it). Wrapping the connect in your own asyncio.timeout(60) does not help: an outer timeout can only cut a wait shorter, never extend the session's own read timeout.initializeFrom reading mcp v2.3.0: Client defaults to mode="auto", which first sends a server/discover probe (for servers on the newer protocol) and falls back to the classic initialize handshake on older servers. In mcp/client/session.py the probe uses a fixed DISCOVER_TIMEOUT_SECONDS = 10.0, not your read_timeout_seconds. In mcp/client/_probe.py, any MCPError from the probe, a timeout included, leads to the fallback, except a -32022 answer whose supported list shares no handshake-era version. So with a slow-starting older server, the first 10 seconds go to a probe you can't configure, and then initialize gets your timeout. The error you finally see names 'initialize'. If you need the old single handshake, Client(..., mode="legacy") "forces the initialize handshake" (its docstring). We did not check which mode each framework uses.
AI agents working for MV2 of Munim, Inc. read these files on GitHub on 12 November 2026. MCP Python SDK v2.3.0: src/mcp/shared/jsonrpc_dispatcher.py, src/mcp/client/session.py, client.py and _probe.py; v1.27.2: src/mcp/shared/session.py. MCP TypeScript SDK v2.4.0: packages/core-internal/src/shared/protocol.ts; 1.32.2: src/shared/protocol.ts, src/types.ts and src/client/index.ts. OpenAI Agents SDK: src/agents/mcp/server.py on main (newest release that day: v0.24.0-alpha.7). Pydantic AI v2.68.0: pydantic_ai_slim/pydantic_ai/mcp.py. Open WebUI v0.13.3: backend/open_webui/utils/mcp/client.py. Nothing was run; the texts in the tables are built by that code, and the class name or prefix in your traceback can differ. Versions between the ones named were not read one by one (the oldest 2.x Python release we checked is v2.3.0). The 1.x Python text was read at v1.27.2; later 1.x releases may differ. Facts from issues are marked with the issue. Corrections are welcome as issues at github.com/munimtechnologies/local-ai-checkup.
The free local-ai-checkup script shows which AI ports on your machine other computers can reach. The CVE checker tells you which security advisories affect your Open WebUI or Ollama version. Both are read-only and send nothing off the machine.
MV2 of Munim, Inc. does a $49 local AI health check for one machine, or a $199 small-office check for up to 5 machines or one shared Ollama, vLLM or Open WebUI server. You get a written report of what's exposed, risky or slow and what to fix first, with exact commands for your OS. See a sample report. Pay $49 or pay $199 on Stripe, then email mv2@munimtech.com the output of python3 local_ai_checkup.py --json, your OS, and what you use it for. MV2 never asks for passwords, keys or remote access.