Microsoft launched Foundry Toolbox and its companion feature Tool Search, a single-endpoint service that lets developers hook AI agents to hundreds of tools without wiring each agent individually. Tool Search cut input tokens by up to 94% when the catalog held more than 600 tools.

Why a central toolbox matters

AI agents need external capabilities—databases, CRMs, analytics platforms—to fulfill user requests. Until now many organisations connected each agent straight to the required APIs. Engineers repeated credential configuration, policy enforcement and error-handling code for every new agent. The result was a tangled web of duplicated settings that was hard to audit and prone to security gaps.

Foundry Toolbox replaces that patchwork with a unified service layer. Instead of a dozen agents each pointing at a dozen separate endpoints, all agents talk to a single “toolbox” endpoint. The toolbox owns versioning, connection strings and security policies, letting teams manage the entire tool ecosystem from one place. Enterprises that run dozens of agents across multiple business units see operational overhead drop immediately.

Large tool catalogs create a hidden cost: token usage. When a language model receives a prompt that lists every available tool, the context window swells, eating up tokens that could otherwise be spent on reasoning or user-facing text. Tool Search attacks the problem at its source.

When an agent enables Tool Search, the model first calls a meta-tool called tool_search, describing in plain English what it needs (e.g., “find the latest sales forecast for region X”). The service returns a short, ranked list of candidate tools that match the intent. The model then invokes call_tool, selecting the most appropriate entry from that list. By exposing only the relevant subset, the prompt stays tiny, saving up to 94% of input tokens in the 600-tool benchmark.

The two-step workflow also improves selection accuracy. In the same benchmark, the model chose the correct tool more often than when it was forced to sift through the full catalog, reducing false calls and unnecessary retries.

How to get the most out of the toolbox

  • Write great metadata – Tool Search relies on each tool’s name and description. Vague labels like “Get data” give the model little to work with. Detailed titles such as “Retrieve customer renewal risks and contacts” guide the search engine to the right match.
  • Pin frequently used tools – If an agent always needs a particular utility on every turn, pin that tool to the agent’s configuration. Pinning skips the search step, shaving latency and token consumption.
  • Organize by capability – Rather than a monolithic toolbox covering the whole enterprise, split tools into logical groups (e.g., sales-tools, CRM-tools). Smaller groups limit the blast radius of misconfigurations and keep search results focused.
  • Test before you deploy – Toolbox versions are immutable; once a version is set as the default, all agents start using it. Use the developer endpoint to validate a new version in isolation before rolling it out company-wide.

These practices matter most when the tool count climbs into the hundreds. For a single agent with a handful of utilities, direct connections may still be the simplest route. But as teams multiply and the toolbox expands, the centralized model pays for itself in reduced duplication, tighter security and measurable token savings.

Bottom line: Foundry Toolbox and Tool Search give large AI deployments a way to tame tool sprawl, cut token waste by up to 94% and enforce consistent security policies—all by adding a single, well-managed service layer. Teams that can afford the initial setup and metadata discipline stand to gain a leaner, more controllable agent ecosystem.