Input aliases

Map client inputs to the inputs a tool expects. AI clients sometimes send mismatched inputs, such as num_results when the tool expects limit, causing the call to fail. The console identifies recurring input mismatches and suggests an alias. Applying it translates the input before policy evaluation and DLP scanning.

How it works

Client sends mismatched input

Sends num_results

Tool expects limit

Call fails ✗

Console suggests a fix

Repeated input mismatch

Shows a quick fix

You apply it

num_results becomes limit

Policies and DLP see limit

Call succeeds ✓

QuilrAI

The banner

When fixes are waiting, the MCP Gateway page shows N quick fixes could prevent failed tool calls - AI clients sent tool inputs that didn't match N times in the last 7 days. Click Review fixes.

Review quick fixes

Show example puts what the client sent next to what the tool expects:

The drawer shows the total at the top, for example 5 quick fixes across 1 server and 26 failed calls in the last 7 days. Fixes are grouped by server, then by tool, with each tool's failure rate, for example get_webpages 22 of 52 calls failed (42%).

Part of a fix card	What it tells you
Headline	What went wrong, for example Clients sent urls - the tool expects ids_or_urls.
Badge	The type of input mismatch (see below).
Failed calls	How many calls failed because of this mismatch.
People	How many users encountered it.
Last seen	When it last happened.
Share of calls	The share of this tool's calls that failed because of this mismatch.
Agent chips	Which AI clients sent mismatched inputs, with their counts.
Show example	A real input as Sent and as Expected, side by side.

Badge	Meaning
required input missing	The client used another name for an input the tool requires.
unknown input	The client sent an input the tool doesn't have, usually an alternative name for an existing input.
inputs in the wrong shape	The client nested its inputs differently from what the tool expects.
value not accepted	The client sent a value the tool doesn't accept, such as another spelling.

Button	Result
Apply	Adds the alias to the server. Subsequent calls use the alias to translate matching inputs.
Dismiss	Hides the suggestion.
Apply all (N)	Applies every fix in the drawer. Shown when there is more than one.

If nothing needs fixing, the drawer says Nothing to fix right now.

Settings > Input aliases

Each server lists its aliases under Configure in the Input aliases section ("Fix inputs AI clients get wrong"). The subtitle shows how many are on, for example 2 of 3 on. A server with aliases also shows an alias count tag on its card.

Part	What it shows
Tool	The tool the alias applies to.
Translation	What the alias does (see below).
Suggested fix / Added manually	Where the alias came from.
Fixed calls	For applied fixes, how many calls it has corrected and when it last did.
Not applying	The tool's inputs changed, so the alias no longer fits and is skipped.
On / Off	Turn the alias off without removing it.
Remove	Delete the alias. Confirm with Remove alias, or click Keep.

The kinds of alias, as the list describes them:

Kind	Example
Renamed input	Translates urls to ids_or_urls
Normalized value	Translates status "open" to "OPEN"
Wrapped inputs	Moves inputs sent at the top level inside arguments
Unwrapped inputs	Moves inputs sent inside arguments to the top level

Aliases are added from Review quick fixes. You cannot add one manually in this section.

Related

• Tool visibility - inspect a tool's input schema.
• Security guardrails - DLP checks run on the translated inputs.
• Allowed Agents - the AI clients named on fix cards.