What is a tool call in MCP?

In MCP (Model Context Protocol), a tool call is a JSON-RPC request the agent sends to a server using the tools/call method. The request names the tool, supplies arguments that match the tool's input schema (returned by tools/list), and the server responds with a structured result the agent treats as grounded data.

How is a tool call different from a REST request?

REST requests are written by human developers using documented endpoints. Tool calls are made by agents at runtime after discovering tool definitions via tools/list, so the agent doesn't need to know the endpoint URL or response shape ahead of time. Tool definitions describe themselves to the agent.

What does Hive's tools/list return?

Hive's root tools/list returns a compact discovery layer, while each category-scoped MCP endpoint returns direct provider tools for that category. The full callable catalog is available through hive://tools and GET /api/v1/tools. Root reads use invoke_api_endpoint, approved Hive state changes use invoke_stateful_endpoint, and category endpoints expose direct tools/call.

Open Dashboard

What is a tool call?

Written by Hive Intelligence, Product docsLast updated July 11, 2026

A tool call is how an AI agent invokes a function: a structured request that names the tool, supplies arguments validated against the tool's schema, and returns a structured response the agent can reason over.

In MCP specifically

The Model Context Protocol formalizes tool calls as a JSON-RPC method:

tools/list returns the catalog of available tools, each with a name, description, and input JSON Schema
tools/call invokes one named tool with validated arguments
The agent never hand-codes a tool wrapper; the schema travels with the tool definition

This is the protocol-level difference between MCP and a raw REST API. With REST, every tool is a custom integration the agent's developer writes. With MCP, every tool is self-describing and discoverable at runtime.

Why tool calls matter for crypto agents

Without tool calls, an agent answering "is this token a honeypot?" has to either rely on its training data (stale, often wrong) or have a developer pre-wire a GoPlus integration, parse the response, and feed it back to the model. With tool calls, the agent calls get_token_security directly with the token address, and Hive's MCP returns a structured grounded answer with fetched_at timestamps.

Hive's release-candidate catalog includes 537 callable tools: 519 wrapped provider tools across 10 categories plus 18 Hive-native stateful tools. Use the appropriate root invoker, a category endpoint, or REST execute after inspecting the schema. See the Live Catalog to confirm deployed availability.

In MCP specifically

Why tool calls matter for crypto agents

Related terms