Tool Use

Structured interface for AI to invoke external functions and APIs

Safely execute LLM-generated code in isolated environments for calculations and data processing

Standardized protocol for sharing context and capabilities between AI models and tools

Exposes a single unified tool interface to the agent while an internal control plane routes each request to the right tool, service, or model behind it. The agent prompt stays stable as the tool catalog evolves, and authentication, policy, rate limiting, and observability are enforced in one place.

Agents operate graphical user interfaces (desktop, browser, or mobile) the way a person does, by reading screenshots and issuing mouse and keyboard actions instead of calling APIs. The core challenge is GUI grounding: mapping an intent to exact pixel coordinates on the screen. The agent runs a loop of screenshot, reason, act, and observe until the task is complete.

Capabilities packaged as self-contained folders. Each skill is a SKILL.md file with YAML frontmatter (name and description) plus optional scripts and resources. Agents discover skills by their metadata and load them through three levels of progressive disclosure: read name and description to judge relevance, load the full SKILL.md once activated, then execute bundled scripts and resources on demand. This keeps context small while exposing many capabilities, and the format is an open standard adopted across vendors.

When an agent can reach hundreds or thousands of tools or MCP servers, loading every schema into the prompt is expensive and hurts selection accuracy. Tool retrieval indexes tool definitions and, for each query, retrieves only the semantically relevant ones (embedding search, active discovery, and reranking) before the model chooses. RAG-MCP reports tool-selection accuracy rising from about 13 percent to 43 percent while cutting prompt tokens by more than half.

Guarantee that model output conforms to a schema by constraining decoding. A JSON Schema, regex, or grammar is compiled into a finite-state machine, and at every decode step the tokens that would violate the schema are masked out so only valid continuations can be sampled. In strict mode this makes malformed JSON and invalid tool arguments impossible. It is available natively at OpenAI, Google, and Anthropic and in open libraries such as Outlines and vLLM.

Instead of emitting one JSON tool call per turn, the agent's action space is executable code that orchestrates tools with loops, conditionals, and variables. A single code block can call many tools, branch on results, and keep intermediate data in the runtime rather than passing every result back through the context window. Anthropic reports roughly a 98 percent token reduction on some multi-tool workflows when tools are called from code via MCP.

A dedicated no-op "think" tool (Anthropic) declared in the agent tool schema that the model calls mid-trajectory to append structured reasoning between a tool result and the next action. The call has no external effect and returns nothing useful: it simply gives the model a checkpoint to re-read policy, verify constraints, and plan before acting inside a long tool-call chain. Anthropic reports a 54% relative improvement on the tau-bench airline pass^1 metric (0.570 versus a 0.370 baseline) when the tool is paired with a prompt that shows how to use it. Distinct from `metacognitive-monitoring`: that entry is the general capability of monitoring one's own reasoning, whereas the think tool is a concrete tool-schema mechanism for it, separate from extended or inference-time thinking that happens before the first response.

A single governed MCP entrypoint that sits in front of many backend MCP servers as a reverse proxy, federating their tools into one curated, managed tool surface that the agent connects to. Every tool call routes through the gateway, which enforces centralized authentication, per-tool authorization that separates the caller's identity from the downstream server's permissions, policy checks, rate limiting, audit logging, and telemetry. This turns a sprawling pile of independently deployed servers into one controlled integration contract, so tools can be curated, versioned, and revoked without touching each agent. Distinct from `control-plane`: the control plane is the general pattern of exposing one interface while routing internally to any tool, service, or model, whereas the MCP gateway is the MCP-ecosystem-specific realization focused on federating many MCP servers and governing the shared tool-access surface.