Logging and Observability

A trace you can read

Record for every call: time, tool name, redacted arguments, duration, success or error, and result size. With these you can answer "why did the agent do that?" and "which tool is slow or failing?". Log to standard error or a file, and redact secrets and personal data before writing.

A logging wrapper

Wrap tool functions so every call is timed and logged. redact removes sensitive values. The log goes to a logger on standard error, never to stdout on a stdio server.

import functools, logging, time
log = logging.getLogger("mcp.tools")

def logged(fn):
    @functools.wraps(fn)
    def wrapper(*args, **kwargs):
        start = time.perf_counter()
        try:
            return fn(*args, **kwargs)
        finally:
            ms = (time.perf_counter() - start) * 1000
            log.info("tool=%s args=%s ms=%.0f", fn.__name__, redact(kwargs), ms)
    return wrapper

Quick check: What must you do to arguments before logging them?

• Nothing, log everything
• Redact secrets and personal data
• Translate them to Hindi
• Encrypt the server