Source: Code execution with MCP (Anthropic, 2025)
Anthropic engineering post: exposing MCP tools as code APIs in a sandbox cut one workflow from about 150k to 2k tokens, at the cost of running a secure sandbox.
Contents
- What: engineering blog post
- Authors: Adam Jones, Conor Kelly (Anthropic)
- Date: 2025-11-04
- Original: Code execution with MCP
Summary
Rather than loading every MCP tool definition into context and routing every intermediate result through the model, present MCP servers as a file tree of code modules the agent imports and calls from code it writes and runs.
Key claims
- One example dropped from about 150,000 tokens to about 2,000 (98.7% saving).
- Progressive disclosure: the agent browses tool files or calls a
search_toolshelper at a chosen detail level. - Intermediate data stays in the execution environment; the model sees only what code logs or returns. The client can tokenize PII before the model sees it and untokenize when passing to another tool.
- Agents can persist state to files and save working code as reusable functions or Skills.
- Caveat: running agent code needs a secure sandbox, resource limits and monitoring, overhead that direct tool calls avoid.