Spot the error in how this MCP server reports a tool failure

Attempt it

Click any words you think contain an error. Click again to unmark.

Mark at least one word to submit.

Quick reference

TL;DR

A failed SQL query is a tool logic error, so it belongs in a successful result with isError true, not in the JSON-RPC error field reserved for protocol faults.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Think of ordering food through a delivery app. The app talking to the restaurant's computer is one layer; the kitchen actually cooking is another. If the app cannot even reach the restaurant, that is a broken connection, a protocol problem. But if the kitchen is simply out of an ingredient, that is not a broken connection at all. The order went through fine. The kitchen just has bad news to report back to you. The MCP server here treats a missing database table like a broken connection. In truth the tool ran perfectly and merely found bad news. That news should ride back inside a normal, successful reply so the model can read it and react, not get thrown away as a transport failure.

Key concepts

isError tool result flag JSON-RPC 2.0 LLM tool use Agent error handling

Quick scan · 2–4 min

Everything important, quickly.

4 min: name the two error classes, place each in the right JSON-RPC field, give the correct isError result shape, and explain the silent failure mode plus the observability cost.

Side-by-side

Aspect	Protocol error (JSON-RPC error)	Tool error (result.isError)
Meaning	Request malformed or method unknown	Tool ran but its logic failed
Envelope	Top-level error object, no result	Successful result with isError true
Example	Parse error, unknown method, invalid params	SQL error, file not found, API timeout
Who consumes it	The host runtime	The model, via the tool result
Model can retry	No, it never sees it	Yes, it reads the message and adapts

Real-world examples

Real products, models, and research that use this idea.

The official MCP TypeScript and Python SDKs return tool exceptions as a result with isError true, reserving JSON-RPC error for dispatch faults.
Claude Code reads isError on each tool result, so a Postgres MCP server's failed query reaches the model and it retries with a corrected query.
The mcp-inspector debugging tool surfaces a tool's isError result distinctly from a transport-level JSON-RPC error.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow should the agent loop behave differently for an isError result versus a JSON-RPC protocol error?

Sign in or go Pro to see the approach for this follow-up.

QWhen IS it correct for an MCP tool call to produce a JSON-RPC error rather than isError?

Sign in or go Pro to see the approach for this follow-up.

QWhat observability problem do you create by collapsing tool errors into protocol errors?

Sign in or go Pro to see the approach for this follow-up.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.