Tool Use and MCP (March 2026)

The Tool-Use Mechanism

Tool use occurs in a 3-step cycle:

1. Schema Presentation: The model is given a JSON schema of the tools.
2. Intent & Extraction: The model outputs a "Call" (e.g., {"tool": "get_weather", "args": {"city": "Tokyo"}}).
3. Execution & Contextualization: The system runs the function and feeds the result back into the prompt.

2025 Nuance: We no longer "hardcode" tool definitions into the system prompt. We use Dynamic Manifests that fetch only necessary tools based on the user's intent.

Model Context Protocol (MCP)

Developed by Anthropic and adopted as an industry standard in 2025, MCP allows models to interact with data and tools regardless of where they live.

• MCP Client: The AI application (e.g., your agent code).
• MCP Server: A standalone process that exposes Tools (Functions), Resources (Data), and Prompts (Templates).
• Communication: Uses JSON-RPC over stdio or HTTP.

Defining High-Precision Tools

A 2025 "Production-Quality" tool must include:

1. Strict Type Validation: Use Pydantic or Zod to enforce schemas before the model even sees the call.
2. Detailed Docstrings: Describe when NOT to use the tool.
3. Confidence Thresholds: Require the model to output a confidence score for the tool call.

# MCP Server Example (Conceptual)
@server.tool()
class ExecuteSQL(PydanticModel):
    """Executes a Read-Only SQL query. DO NOT use for DROP/DELETE."""
    query: str = Field(..., description="The SELECT query to run.")

    async def run(self):
        # Implementation here...
        pass

MCP vs. OpenAI Function Calling

Streaming Tool Calls

Frontier models support Partial Tool Speculation. Instead of waiting for the full JSON to generate, the system starts "prefetching" tool results as soon as the tool name and critical IDs are visible in the stream. This reduces perceived latency by 400-800ms.

MCP 2.0: Streamable HTTP & Auth (2026)

The MCP 2.0 specification (ratified March 2026) introduced two major changes:

[MCP Client] ←── Streamable HTTP POST /mcp ──→ [MCP Server]
                  (with SSE response stream)

{
  "type": "oauth2",
  "grant_type": "client_credentials",
  "scopes": ["tools:read", "resources:documents"]
}

Computer-Use Tools (Anthropic)

Claude 3.5+ introduced native computer-use tools — the model can directly control a desktop or web browser. These are available via the Anthropic API:

import anthropic

client = anthropic.Anthropic()

response = client.beta.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=4096,
    tools=[
        {"type": "bash_20250124", "name": "bash"},
        {"type": "text_editor_20250124", "name": "str_replace_based_edit_tool"},
        {"type": "computer_20251022", "name": "computer",
         "display_width_px": 1280, "display_height_px": 800}
    ],
    messages=[{"role": "user", "content": "Open Firefox, go to GitHub, and clone my repo."}],
    betas=["computer-use-2024-10-22", "interleaved-thinking-2025-05-14"]
)

Production safety rules for computer-use:

1. Always run in a sandboxed VM (Docker + VNC, or E2B cloud)
2. Screenshot-validate critical state before destructive actions
3. Use HITL (Human-in-the-Loop) for irreversible actions (file deletion, form submission)
4. Set ANTHROPIC_MAX_COMPUTER_TOKENS to cap runaway loops

Context7: Live Documentation MCP

One of the most practical MCP servers in 2026 is Context7 — it resolves the "stale training data" problem for coding agents:

# Without Context7:
Agent: "I'll use langchain's `create_openai_tools_agent` function..."
(This function was deprecated 6 months ago)

# With Context7 MCP:
Agent → MCP: list_resources("langchain")
MCP → Agent: Returns current v0.3.x docs
Agent: "I'll use the new `create_react_agent` interface..."

Setup in Claude Desktop / Claude Code:

{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp"]
    }
  }
}

Claude automatically calls resolve-library-id and get-library-docs before writing code that uses the library.

Interview Questions

References

• Anthropic. "The Model Context Protocol Specification" (2025)
• JSON-RPC 2.0 Specification.
• Pydantic v3.0 Documentation.

Next: Multi-Agent Orchestration