loupe-ai 0.0.72

A magnifying glass for your AI agent — open-source forensics + interpretability for LLM agents.

loupe — Python SDK

Forensic observability + circuit attribution for Python LLM agents.

Install

pip install -e '.[ui]'           # from this repo (canonical install today)
pip install -e '.[interp]'       # adds the real SAE attribution backend
pip install -e '.[ui,interp,langgraph,anthropic,openai,universal]'

Quickstart

from loupe import trace, record_step
from loupe.integrations import patch_all

patch_all()                                  # auto-capture any installed LLM SDK

@trace(framework="anthropic")
async def my_agent(query: str):
    record_step("plan", "compose request")
    # ...your real agent code; LLM calls captured automatically
    return result

await my_agent("refactor auth.py")
# trace saved to ~/.loupe/traces/{run_id}.jsonl

View traces locally:

loupe ui              # opens http://localhost:7860 — live SSE dashboard
loupe list            # terminal table of every run
loupe attribute <id>  # SAE circuit attribution per llm-call step
loupe cluster         # find features that recur across tagged failures
loupe replay <id>     # re-invoke a captured run for reproducibility testing

See SPEC.md for the wire format and ARCHITECTURE.md for the layering.