Python SDK

Trace, replay, and cluster your LLM agent runs. Wraps OpenAI, Anthropic, LangChain, LlamaIndex, and MCP.

Pre-release. The SDK is not published yet.

The package below is not on a public registry yet, so the install command will not work today. The API shown here is a preview of the shipping interface. Want early access? Sign in to request it.

Install

Once the package is published, install with pip:

pip install verdict-trace

Optional per-provider integrations:

pip install "verdict-trace[anthropic]"
pip install "verdict-trace[openai]"
pip install "verdict-trace[langchain]"

Configure

import verdict
from verdict_trace import Settings

verdict.configure(Settings.from_env())
verdict.instrument(anthropic=True, openai=True, langchain=True)

Environment variables:

VERDICT_API_KEY (required, starts with vdt_)
VERDICT_BASE_URL (ingest API, https://verdict-api.jaitly.cloud)
VERDICT_BATCH_SIZE (default 50)
VERDICT_FLUSH_MS (default 200)
VERDICT_BUFFER_DIR (default ~/.verdict/buffer)

Trace an agent run

async with verdict.trace("my-agent") as t:
    response = client.messages.create(...)
    t.add_step(type="decision", payload={"chose": "tool_x"})

Any instrumented Anthropic, OpenAI, or LangChain call inside the trace block is captured automatically. View runs at verdict.jaitly.dev.

Guarantees

Non-blocking: steps buffered, flushed every 200ms or 50 steps.
Retry: 5 attempts, exponential backoff 1 → 2 → 4 → 8 → 16s.
Dead-letter recovery: failed batches saved to ~/.verdict/buffer/, replay via client.recover_deadletter().
Idempotent: step ids are client-generated (UUID7); the server dedupes.