What is 'tool use' in the context of LLMs?

'Tool use' is Anthropic's terminology for allowing a large language model (like Claude) to interact with external systems. The model responds with a 'tool_use' block, pausing text generation while the host application executes the tool and returns a 'tool_result'.

How does tool use differ from OpenAI's function calling?

While conceptually similar, the JSON schema differs. Anthropic uses 'tool_use' and 'tool_result' block types embedded in a single content array, whereas OpenAI historically used dedicated 'function_call' and 'tool_calls' attributes on the message object.

How important is latency for tool use?

Extremely important. An agentic loop may require multiple sequential tool calls. If each call takes 1 second, a 5-step loop takes 5 seconds of dead time. Platforms like wmcp.sh provide sub-100ms execution to keep reasoning loops fast.

How do I secure credentials during tool use?

Never include raw API keys in the tool schema sent to the model. An encrypted credentials vault should intercept the tool request server-side and inject the necessary authentication before hitting the target system.

GLOSSARY · /GLOSSARY/TOOL-USE

Tool Use (Anthropic)

When a language model hits the limit of its static training data, it needs to look up facts, run calculations, or mutate state. "Tool use" is the mechanism that allows it to pause generation and request execution, driving modern AI workflows.

the gap

Why agents need specialized execution environments

In Anthropic's ecosystem, when Claude determines it needs external data—for instance, to fetch a customer profile for support@example.com at Acme Corp—it emits a tool_use block instead of standard text. The client application must halt, execute the requested tool, and append a tool_result back to the message history.

However, this agentic loop is fragile. If the host environment is slow, the model waits. If authentication fails, the context window fills with error tracebacks. wmcp.sh abstracts this execution layer, ensuring that when Claude asks for data, it receives a normalized, sub-100ms response. (Note: wmcp.sh is not affiliated with Anthropic, OpenAI, or any model provider.)

the architecture

How tool use looks in practice

import Anthropic from "@anthropic-ai/sdk";

const anthropic = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });

// Fetch pre-formatted tool schemas from wmcp.sh
const toolSchemas = await fetch("https://api.wmcp.sh/v1/tools?url=...").then(r => r.json());

const msg = await anthropic.messages.create({
  model: "claude-3-7-sonnet-20250219",
  max_tokens: 1024,
  tools: toolSchemas,
  messages: [{ role: "user", content: "What is Acme Corp's MRR?" }]
});

// Model returns a tool_use block
if (msg.stop_reason === "tool_use") {
  const toolBlock = msg.content.find(c => c.type === "tool_use");
  console.log("Executing tool:", toolBlock.name, toolBlock.input);
  // wmcp.sh executes this securely via its credentials vault
}

capability

Optimizing Tool Use Loops

Capability	Standard Architecture	With wmcp.sh
Execution Latency	⚠️ 500ms - 2s typical for serverless stacks.	✅ Sub-100ms at the edge.
Caching Strategy	❌ Non-existent or manual Redis wiring.	✅ Built-in short TTL (~1s) micro-caching.
Credential Safety	❌ API keys passed directly in `tool_use` input.	✅ Encrypted credentials vault handles keys.
Schema Alignment	⚠️ Needs manual translation from OpenAPI to Anthropic JSON.	✅ Zero-config schema bridging.

FAQ

Common questions.

Is tool use the same as function calling?

Conceptually yes. "Tool use" is Anthropic's terminology, while "function calling" is traditionally associated with OpenAI. Their JSON payload structures differ, but both enable the LLM to trigger external code. wmcp.sh normalizes both.

Why does latency matter for tool use?

A complex agentic task may require 10 sequential tool calls. If your tools take 2 seconds each to resolve, the user stares at a spinner for 20 seconds. Sub-100ms latency ensures the reasoning loop remains fluid and real-time.

Can a model call multiple tools at once?

Yes, Claude supports parallel tool use. It can emit multiple tool_use blocks in a single turn, which wmcp.sh can then execute concurrently to drastically reduce total time-to-completion.

Need this done for you?

Skip the wiring — we build, deploy, and monitor.

Custom adapter + hosted MCP at mcp.yourbrand.com + verified badge. Pricing: Starter $499 one-time, Managed Retainer $999/mo, or Enterprise $4,999+/mo.

See /managed → Submit (free)