spanix

A local profiler for AI agents. One line, no account, no server.

An agent SDK reports cost, tokens and duration on every message it streams. It reports them exactly once, and the caller's for loop throws them away. spanix stops throwing them away.

from claude_agent_sdk import query
from spanix import watch, last_run

stream = query(prompt=p, options=opts)

async for msg in watch(stream, run="nightly"):
    handle(msg)

print(last_run().summary())

spanix · nightly  22.6s
  cost      $0.6512
  tokens    206,431 (in 184,220 · out 8,911 · cache 13,300)
  turns     11
  messages  38
  tools     Read ×1, WebFetch ×5, Grep ×1

The message is yielded untouched, so the loop around it does not change.

Design

It never breaks your agent. Every recording path is wrapped in its own try, and the message is always yielded. A profiler that takes down production is a profiler nobody installs.

It never imports your agent SDK. Fields are read by name with fallbacks, so pip install spanix pulls in nothing, and supporting a second runtime is just another list of names. There are zero dependencies.

It never keeps content. Prompts, tool arguments and tool results carry API keys and customer data. Only counters and tool names are stored.

Status

0.0.1 is the accounting core: watch(), in-memory Run records and summary(). Nothing is written to disk and nothing is sent anywhere.

Next: SQLite at ~/.spanix/runs.db, the subagent tree with cost rolled up per node, and a panel on localhost.

License

Apache-2.0