Give your agent only the tools each turn needs.
Toolcue picks the right tools for every user message, including the lookups they depend on, in tens of milliseconds before your LLM call. Smaller prompts, fewer wrong tool calls, and a catalog that can keep growing.
Works with any model and any framework: you keep your agent and your tools.
- refund_invoice MATCH
- get_customer_invoices LOOKUP supplies the invoice id the refund needs
- find_customer LOOKUP supplies the customer id the invoice list needs
- create_invoice CLOSE MATCH
- cancel_subscription CLOSE MATCH
- ask_user ALWAYS ON
Every tool in every prompt doesn't scale.
Most agents send their whole tool list on every turn. A "thanks!" costs as much as a refund, the model has more ways to pick the wrong tool, and the 41st tool makes every turn slower and more expensive.
Tokens on every turn
Tool definitions are often most of the prompt. You pay for all of them on every call, including the ones the turn could never use.
Wrong or missing tool calls
Look-alike tools compete for the same request, and multi-step tasks fail when the model doesn't know it needs a lookup first.
A ceiling on growth
Each new tool makes the problem worse, so teams stop adding tools, or split one agent into many just to keep prompts small.
One call before your LLM call.
Publish your tools once. On each user message, ask Toolcue which ones this turn needs, and pass your model only those.
Publish your tools
Send your tool definitions as they are: OpenAI, Anthropic and MCP formats convert in one line. Publish again whenever they change.
Ask per turn
Send the user's message and the recent conversation. Toolcue answers with tool names and the reason for each.
Call your model
Pass only those tools. If the
model still needs another, a small find_tools tool recovers it in the same turn.
It understands the task, not just the words.
Matching a message to a tool description is the easy part. Toolcue also knows how your tools depend on each other and what already happened in the conversation.
How it works in the guide- The lookups a task depends on"Refund Globex's last invoice" also gets the invoice list and the customer search, because the refund needs an invoice id and the list needs a customer id.
- Ids already in the conversationWhen an earlier tool returned
customer_id: cus_8812, Toolcue hands it back, so the model doesn't look the customer up twice. - Follow-ups and small talk"ok, move them to the scale plan" works with the history; "thanks!" gets only your always-on tools.
- RolesAdmin-only tools stay hidden from everyone else, in selection and in search.
- Any languageUsers ask in their own words and language; your tool descriptions stay as they are.
- No LLM insideSelection runs in tens of milliseconds with no model call of its own, so it adds almost nothing to your response time.
Measured on real conversations.
Production conversations of a B2B SaaS assistant, held out from tuning, plus a synthetic test where the catalog grows by 60 tools it has never seen used.
94–97% right after the catalog grew by 60 new tools.
The bigger the catalog, the bigger the saving.
Median, ~55 ms at p95, measured on a laptop. Repeated messages are faster.
How we measured, and what we compared against: benchmark.
Keep your stack.
Toolcue returns tool names. Your agent, your model and your tool code stay exactly where they are. Tested integrations for:
Join the waitlist.
We're onboarding teams in small groups. Tell us a little about your agent, and we'll send you an endpoint and keys when your access is ready.
Best fit: agents with 20 or more tools, or catalogs that keep growing.
Questions
Does Toolcue call an LLM?
No. Selection runs without any model call, which is why it takes tens of milliseconds and costs no tokens.
What if it leaves out a tool the model needs?
Give your model a find_tools tool:
it searches your whole catalog and adds what it finds in the same turn. And if Toolcue is ever slow or
unreachable, send all your tools for that turn.
How many tools do I need?
Under about 15–20 tools, sending all of them is fine. Toolcue pays off as your catalog grows, and keeps paying off when it grows by dozens of tools a month.
Which models and frameworks?
Any. Toolcue returns tool names; you pass the matching definitions to whatever model and framework you use.
Can we run it ourselves?
A self-hosted image is planned for teams that need to keep everything in their own cloud. Tick the box on the waitlist form.
Who is behind Toolcue?
Toolcue is built by Foldspace, which builds in-app AI agents for software products.