llmwho

Passive LLM endpoint identity and stability instrumentation for Python.

pip install llmwho

import llmwho

handle = llmwho.init()

Direct Anthropic Messages API calls need no LLMWho-specific wrapper:

import llmwho
from anthropic import Anthropic

llmwho.init()
client = Anthropic()
client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=64,
    messages=[{"role": "user", "content": "Hello"}],
)

One call instruments installed HTTPX (sync and async) and requests clients for recognized LLM routes. It writes content-free observations to ~/.llmwho/events.jsonl; it never sends active traffic. Calls are idempotent, streaming responses are not consumed, and handle.shutdown() safely restores only LLMWho-owned patches.

Send the same observations to a self-hosted Collector without changing request sites:

handle = llmwho.init(
    collector_url="http://127.0.0.1:7734",
    collector_token="…",
)

The remote queue is bounded, non-blocking, and fail-open. Native and OTLP modes are available; no Collector URL keeps the local JSONL default. Run the service with llmwho collector; deployment and security details are in the Collector guide.

The CLI also accepts content-free Claude Code and Codex lifecycle events: llmwho hook claude-code|codex --event EVENT. Project configuration and privacy details are in the agent hook guide.

Run an explicit OpenAI-compatible compatibility canary:

report = llmwho.probe(
    base_url="https://api.openai.com/v1",
    api_key="…",
    model="gpt-4o-mini",
)

Compare explicitly supplied output corpora without network or persistence:

report = llmwho.science.output_affinity_matrix({
    "reference": ["reference answer"],
    "endpoint": ["endpoint answer"],
})
print(report["evidence"]["matrix"][0][1])

The symmetric character-trigram distance measures surface style only. It does not prove model identity, distillation, or capability transfer. Full method and limits: Output-affinity matrix.

Inspect local history:

llmwho summary
llmwho dashboard

Version 0.4 records the response body's declared model field only as a provider declaration. It never turns that declaration into model identity or confidence. Identity remains unknown unless an independent calibrated detector supplies evidence. Undocumented model response headers are ignored; the smoke suite measures capability and does not uniquely identify arbitrary model weights. Raw prompts, responses, headers, query strings, and credentials are never stored.

Full documentation, research review, and source: github.com/tcztzy/llmwho.