Tool lists

Send AI apps a short tool list: a server's most-used tools in full, with every other tool one search away.

Every turn, an AI app sends the model the full description of every tool on every connected server, including tools nobody uses. A server with 60 tools can add 8,500 tokens to every turn even when people only ever call three of them. Tool lists cut that down per server:

Switch	What AI apps get
List most-used tools	The tools your people use most, listed in full, plus find_relevant_tools (finds any other tool) and call_tool (runs a found tool).
Add Code mode	run_code: the AI writes a short script that calls the server's tools and returns only the answer, instead of calling tools one by one and reading every result.

Both switches are off by default. With them off, AI apps get every tool, exactly as before.

Nothing is blocked

A short list only changes what is listed up front. Every tool stays callable by name, through call_tool or from a run_code script, and every call runs through the same tool permissions, guardrails, approvals, quotas and logs as a direct call.

How it works

Measure

Tool calls from the last 30 days

Ranked hourly, per server

Listed in full

Tools that cover about 90% of calls

3 to 15 tools, plus pins

One search away

find_relevant_tools finds the rest

call_tool runs it

Still governed

✓ Same rules, guardrails and logs

✓ Nothing blocked

QuilrAI

• Which tools are listed. The gateway ranks each server's tools by how often your people called them in the last 30 days (recent calls count more) and lists the few that cover about 90% of calls, between 3 and 15 tools. Pinned tools are always listed. The ranking refreshes every hour, so the list follows how people actually work.
• Small lists stay as they are. If a server's full list is already short, the gateway sends it unchanged.
• Stable during a chat. An AI app picks up a changed list when it starts a new chat.

Where the short list applies

Where	With List most-used tools on
The server's own address (<YOUR_MCP_BASE_URL>/<server-slug>/mcp)	The server sends its most-used tools, find_relevant_tools and call_tool. With Add Code mode on too, it also sends run_code.
OneMCP with dynamic tool calling on	The server's most-used tools are listed directly next to OneMCP's own tools, named <server-slug>_<tool> (for example mcp-query_llm_alert_scan). Its other tools stay reachable through find_relevant_tools and call_tool. run_code is offered when any server the person can use has Add Code mode on.
Workflow Agents	Not affected. An agent's tools are chosen when it is built, so agents always get them in full.
Local MCP servers	Not affected. Local servers always send their full list, and the setting is not shown for them.

With both switches off on every server, OneMCP behaves exactly as described in Smart Tools.

Configure it on the server

Go toQuilrAI consoleSettingsAI GatewayMCP Gatewayserver cardConfigureTool list

The What AI apps get card holds the two switches. Turn on the ones you want, then click Save tool list in the footer. The section list shows the result: Off, Most-used tools, Most-used + Code or Code only.

Setting	What it does
List most-used tools	Sends the server's most-used tools in full and keeps every other tool one search away, on the server's own address and in OneMCP sessions with dynamic tool calling on.
Add Code mode	Lets AI apps run short run_code scripts that call this server's tools and return only the answer. Best for data-heavy servers. On the server's own address it comes with the most-used tools list.
Pin	Always list a tool in full, even when it is not among the most used. Pin from the tools one search away list or the Most often found by search list. Pinned tools show with a pin; click the x on a pinned tool to unpin it.

Preview

The Preview card shows the short list this server sends with List most-used tools on, based on how your people used it recently. It works while the switch is off, so you can see the effect before you turn it on.

Part	What it shows
Tokens per turn	The full list against the short list, with the cut (for example 8.5k → 782, -91%).
Coverage	How many listed tools cover what share of your people's calls on this server.
Listed in full	The most-used and pinned tools, plus the tools the gateway adds (find_relevant_tools, call_tool, and run_code with Code mode).
Tools one search away	Every other tool, with its share of recent calls and a Pin button.
Refresh usage	Re-reads usage now instead of waiting for the hourly refresh.

Once the server has served short lists, a What the short list saved here card adds the tokens not sent, the estimated cost avoided, the share of calls that went through search (under 10% means the listed tools fit the work; above it, consider pinning) and, with Code mode, the scripts run.

Code mode (run_code)

run_code runs a short Python script in a sandbox. The script calls tools with await call_tool(name, args) and returns only what the AI needs, so large intermediate results never reach the model.

Limit	Value
Run time	60 seconds
Memory	64 MiB
Tool calls per script	50
Files, network, imports	Not available

Each tool call inside a script is a normal governed call: policies, approvals, guardrails and logs apply to it as if the AI had called the tool itself. If a script stops early, the result lists any tool calls that were still running and whether each one finished, so the AI does not repeat work by mistake.

Find servers worth turning on

The MCP Gateway page shows a Savings opportunity card when servers send many tool descriptions nobody uses.

Click Review to open Send AI apps shorter tool lists. Each server shows its tokens per turn today and with the short list, the share of calls the listed tools cover, and the tokens a month you could stop sending. Open settings opens the server's Tool list section, and Turn on turns List most-used tools on for that server, keeping its pins and Code mode setting.

A server is recommended when the short list saves at least 2,000 tokens per turn and the server had at least 50 calls in the last 30 days. When the Policy Engine is on, the drawer shows a note instead of Turn on (see Going further with the Policy Engine).

Where savings show

Place	What you see
Server card	A Tool list chip once a switch is on, for example Most-used · -91% or Most-used · -91% + Code.
Server settings > Tool list	The preview and, once used, What the short list saved here.
Overall analytics > Servers & tools	The Tool list savings panel, below.

Column	Meaning
Tool list	The server's switches (Off, Most-used tools, Most-used + Code, Code only); More to save marks a recommended server.
With a short list	Share of the server's calls in the period made with a short list.
Tokens per turn	Full list against short list, with the cut.
Est. saved (30 days)	Tokens and cost not sent, or / month if turned on for a server that is off.
Through search	Share of short-list calls that went through find_relevant_tools and call_tool. pin more means the listed tools miss common work; pin the tools found most often.

Measured and estimated

The savings figures are estimates built on real usage. They are not read from a model provider's bill.

Part	Source
Which calls were made with a short list, and through search or run_code	Measured: counted from the gateway's tool call logs.
Tool list size, full and short	Measured from the server's real tool descriptions, then converted to tokens at about four characters per token.
Tokens not sent	Estimated: the per-turn cut times the calls made with a short list, treating each tool call as one model turn. A chat that makes several calls in one turn, or many turns without calls, saves more or less than this.
Cost avoided	Estimated at a reference rate of $3 per million input tokens, not your provider's price or any caching it applies.
Recommended and "if turned on" figures	Projected from the last 30 days of calls made with the full list.

Tool list savings show only in the places above. The Tokens saved figures on the MCP Gateway page and in Costs & Savings count token saving strategies.

Going further with the Policy Engine

When the Policy Engine is on for the MCP Gateway, this section turns read-only and the OneMCP Features card (stage 1, Session) in Govern > Policy Engine > MCP Gateway applies instead. Edit anyway changes the stored values, which are used only if the Policy Engine is disabled (see What happens to classic settings). Pins always come from the server settings.

The card has two effects for tool lists, both Off for everyone until a configuration turns them on:

Effect	Card row	Merge rule
tool_list.top_tools	Most-used tools listed	Highest priority wins; at equal priority on wins.
tool_list.code	Scripts (run_code)	Highest priority wins; at equal priority on wins.

To turn tool lists on for a server, click Quick add under Most-used tools listed, choose MCP server under Applies to and pick the server, set Most-used tools listed (and Scripts (run_code) if you want scripts) to On, then Add to draft and Review & publish.

Scenarios the card supports that server settings cannot:

• Short lists for some people only. Match smart groups or user email, for example turn short lists on for a pilot group first.
• Scripts on one route. Match route kind to offer run_code through OneMCP but not on a server's own address.
• Many servers at once. Match MCP name with a list or pattern instead of configuring each server.

short_list_for_reporting_serverssession

runs on session·priority 650

WhenMCP nameis any ofmcp-queryQuilr SE-CS MCP

Then

list most-used tools→true

Related

• OneMCP - one endpoint with dynamic tool calling and smart tools.
• Token saving - shorten tool results; tool lists shorten what is sent before any call.
• Tool visibility - hide tools from AI apps entirely.
• Group and user rules - per-group overrides for server settings.
• Policy Engine overview - how cards, stages and priorities work.