Judgeval AI Agent

<div align="center"> <a href="https://judgmentlabs.ai/"> <picture> <source media="(prefers-color-scheme: dark)" srcset="assets/logo_darkmode.svg"> <img src="assets/logo_lightmode.svg" alt="Judgment Logo" width="400" /> </picture> </a> <br>

The Continuous-Improvement Stack for Agents

Detect failures, triage root causes, and ship fixes backed by production data.

</div>

Overview

Judgeval is an open-source Python SDK for agent improvement. It provides tracing and Agent judge evaluation for LLM-powered applications — so you can detect failures, understand what went wrong, and validate fixes against real production cases before shipping.

To get started, dive into the docs.

Why Judgeval

OpenTelemetry-based tracing -- Instrument any function with @Tracer.observe(). Automatically captures inputs, outputs, and LLM token usage. Built on OpenTelemetry for full compatibility with existing observability stacks.

Agent judges -- Define prompt-based scorers to evaluate traces at scale. Judges apply labels to traces — a binary judge applies its label when true, and a classification judge applies the chosen option — and those labels accumulate into a searchable record of how your agent acted over time. Run judges against live production traffic or replay them on historical traces to validate fixes before shipping.

Online monitoring -- Automatically score live production traffic server-side with no latency impact. When a judge applies a label, that label surfaces as a structured signal — configure Slack alerts so regressions and recurrences never go unnoticed.

Broad integrations -- Auto-instrumentation for OpenAI, Anthropic, Google GenAI, and Together AI. Framework support for LangGraph, and Claude Agent SDK.

Quickstart

Install the SDK:

pip install judgeval

Set your credentials:

export JUDGMENT_API_KEY=...
export JUDGMENT_ORG_ID=...

Add observability to your agent with two lines of setup:

from judgeval import Tracer, wrap
from openai import OpenAI

Tracer.init(project_name="my-project")
client = wrap(OpenAI())

@Tracer.observe(span_type="tool")
def search(query: str) -> str:
    results = vector_db.search(query)
    return results

@Tracer.observe(span_type="agent")
def run_agent(question: str) -> str:
    context = search(question)
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": f"{context}\n\n{question}"}],
    )
    return response.choices[0].message.content

run_agent("What is the capital of the United States?")

SQL

Use Judgeval.sql(sql_text) for read-only queries against Judgment's virtual schema, which abstracts the underlying storage. The server validates incoming queries, rejects writes, and enforces organization and project scope. The client uses its existing API key, organization membership, and resolved project. Viewer access and the public query rate limit apply.

from judgeval import Judgeval

client = Judgeval(project_name="my-project")
print(client.discover_schema())
result = client.sql("SELECT count() AS run_count FROM telemetry.traces")
print(result["rows"])

client.discover_schema() returns a Markdown string with the server's generated tables, column types and descriptions, row semantics, examples, and query limits, using the same reference as MCP discover_schema. It contains no project data and requires organization viewer access, but no resolved project. The HTTP equivalent is GET /v1/sql/schema, which returns {"schema": "...Markdown reference..."}.

For query execution, use POST /v1/projects/{projectId}/sql with Authorization: Bearer <api-key>, X-Organization-Id: <organization-id>, and JSON body {"sql": "SELECT count() AS run_count FROM telemetry.traces"}. Organization and project scope are derived by the server. Use SQL predicates on supported catalog columns and LIMIT to narrow results. Physical tables, writes, multiple statements, and caller-specified execution limits are unsupported. DAL catalog allowlists, tenant isolation, and result limits of 1,000 rows and 5 MiB apply; over-limit results return an error. SQL text must contain a non-whitespace character and cannot exceed 50,000 characters.

The response contains catalog_version, columns (name, type, nullable), rows, row_count, and elapsed_ms. SQL integers outside JavaScript's safe range (-(2**53 - 1) to 2**53 - 1) arrive as exact decimal strings, including inside nested arrays and objects. For example, 9007199254740993 arrives as "9007199254740993"; use int(value) when you need a Python integer. Small integers and floating-point values remain numbers, and column types retain their original SQL types. Validation and execution errors use the SDK's existing exception mapping.

Integrations

Supports OpenAI, Anthropic, Google GenAI, Together AI, LangGraph, and Claude Agent SDK. See the full integrations docs.

CLI

Manage agents, traces, judges, labels, and evaluations from the terminal. Query trace history, deploy judges, review the labels judges applied, and run evals against production data — all without leaving your shell. See the CLI repo and docs.

MCP Server

Connect Judgment to any MCP-compatible AI tool. Query agent traces, invoke judges, browse the labels judges applied, and surface failures directly inside your AI assistant or IDE. See the docs.

Links

• Documentation

---

Judgeval is created and maintained by Judgment Labs.