Repro: widget tool result is dropped when the handler loses a race with the iframe load

In mcp-use dev, a tools/call that renders a widget can hang in the inspector: the server finishes the call, but the result never reaches the client, which spins until its request timeout (-32001). Whether it hangs depends on how long the tool handler takes — it's a probabilistic race, not a fixed timeout.

Measured on the stock scaffold by sweeping the handler delay (DELAY_MS):

The probability of hanging climbs smoothly from ~150ms to ~300ms — a transition band, not a cliff. Reading: the tool result is racing the widget iframe finishing its load/connect (itself a variable ~150–300ms). If the handler returns before the iframe is done, the result is delivered; if it returns after, the result is dropped.

This is why it looks intermittent in a real server: a handler that hits an API/DB has jittery latency that lands in the transition band, so it hangs only some fraction of the time.

• Only in mcp-use dev. A production build / deployed connector does not hang.

• This project is the unmodified create-mcp-use-app --template mcp-apps scaffold; the only change is making the stock search-tools 2s delay configurable via DELAY_MS.

Environment

• mcp-use: 1.33.0 (also seen on 1.30.0). macOS, Chrome.

Related MCP server: mcp-elicitation-proxy

Getting Started

npm install
npm run dev

Steps

DELAY_MS=2000 npm run dev      # http://localhost:3000, opens the inspector

1. Open the inspector, call search-tools (args can be empty {}) → spins forever, eventually -32001. (At 2000ms it hangs every time.)

2. Restart with DELAY_MS=0 npm run dev, call again → delivers instantly, every time.

3. To see the race, sweep DELAY_MS=150 / 200 / 250 (restart each time) and call the tool several times — it hangs a growing fraction of the time as the delay rises.

On hanging runs the widget iframe renders its skeleton (it received the tool input) but never the output (props).

Negative controls — Node clients do NOT reproduce it (so it's browser-specific)

These are MCP clients, so the dev server must already be running in another terminal. DELAY_MS is a server setting (it controls the handler) — set it on npm run dev, not on the script. Two terminals:

# terminal 1 — start the server at the hang-inducing delay, leave it running
DELAY_MS=2000 npm run dev

# terminal 2 — run a client against it
node repro.mjs          # SDK client: pre-read widget, then tools/call + resources/read CONCURRENTLY
node repro-flood.mjs    # SDK client: tools/call + a burst of widget asset fetches, concurrently

(If you see ECONNREFUSED ::1:3000, the dev server isn't running — start terminal 1 first.)

Both deliver the tool result fine even at DELAY_MS=2000 (the browser hangs every time at that delay). A headless mcp-use client screenshot --tool search-tools also works. These never load the widget iframe in a browser, so there's nothing for the result to race — consistent with the race reading above.

Why (unconfirmed)

Browser-only + dev-only, probabilistic in a ~150–300ms band → looks like the result racing something that completes when the widget iframe loads/connects, rather than a fixed timer or plain connection contention. In mcp-use dev the iframe's assets are served from the same origin (localhost:3000) by Vite; a deployed build serves prebuilt assets from a separate CDN origin and doesn't hang. Exact mechanism unconfirmed.