Model Context Protocol
Model Context Protocol
The Model Context Protocol (MCP) is a standardized protocol developed by Anthropic for connecting AI models and agents to external tools through structured, typed tool definitions. MCP provides type-safe schemas and discoverable tool catalogs, enabling agents to invoke tools through their hosting framework’s native tool-calling interface rather than parsing CLI output.
How It Works
An MCP server exposes a set of tools, each defined with a JSON schema specifying its name, description, and parameter types. The agent’s host framework loads these schemas into the model’s context, and the model invokes tools by generating structured calls that match the schema. Results are returned as structured data.
This contrasts with CLI-based approaches where the agent runs shell commands and parses text output. See Agent-Tool Interfaces for a full comparison.
Strengths
- Type safety — tool parameters and return types are explicitly defined, reducing misuse
- Discoverability — tool catalogs are programmatically enumerable
- Standardization — a single protocol works across different hosting frameworks and models
- Ecosystem — growing library of MCP servers for common services (GitHub, databases, browsers, cloud providers)
Limitations
Benchmark research has quantified several costs:
- Schema overhead scales with tool count. A browser MCP server exposing ~30 tools inflates input tokens to 185K per task vs. 79K for an equivalent AXI CLI — a 2.3x overhead that compounds across every turn.
- Poor discoverability in practice. Lazy loading leads to tool name confusion — agents select
take_screenshot(1MB base64 PNG) instead oftake_snapshot(80KB text) and crash sessions. Eager loading inflates context even further. - No shell composability. MCP results pass through the model’s context window. Agents cannot pipe output through
greporheadto filter before it consumes tokens. - Separate action and observation. Like raw CLIs, MCP tools often require separate calls for action (click, navigate) and observation (snapshot), doubling turn count.
Atlassian’s MCP Compressor demonstrates that wrapping MCP servers as CLI subcommands recovers most of the efficiency — achieving 100% success at $0.091/task vs MCP’s $0.100, by eliminating schema overhead and enabling shell composability.