---
title: "Use Judgment from a coding agent"
description: "Combine Judgment project data and workflow guidance in a coding agent without confusing their roles."
sidebar:
  label: "Use Judgment from a coding agent"
seo:
  title: "Use Judgment from a Coding Agent"
  description: "Choose between the Judgment MCP server, repo-local skill, and in-product Judgment Agent."
---

Give a coding agent the smallest set of Judgment surfaces that fits the job.
For production debugging and code changes, connect the Judgment MCP server and
install the repo-local skill. Use Judgment Agent when page context and selected
entities matter.

## Choose the right surface

| Surface | What it supplies | Best for | Boundary |
| --- | --- | --- | --- |
| **Judgment MCP server** | Authenticated organization and project data, documentation search, and read and write tools | Investigating real traces, judges, datasets, tests, prompts, issues, and automations from a coding agent | Writes may require a role or client approval |
| **Repo-local Judgment skill** | Versioned instructions that teach an agent when and how to use Judgment | Applying repeatable tracing and investigation practices while editing a repository | Guidance only; it does not grant credentials or add tools |
| **Judgment Agent** | Judgment data tools plus the current product page, selections, and conversation context | Investigating and acting from inside the Judgment platform | In-product context; it does not read or edit your local repository |

The Judgment MCP server includes `search_docs` and `read_doc_page`, so the same
connection can retrieve current product and SDK guidance.

## 1. Connect the Judgment MCP server

Follow [MCP setup](/documentation/mcp-and-agent-tools) to connect your coding
agent and verify access to the intended organization and project.

## 2. Install the repo-local skill

Follow [Install the Judgment agent
skill](/documentation/mcp-and-agent-tools/judgment-agent-skill) from the root of
the repository the agent will edit. A skill improves tool selection and
workflow, but the data MCP connection remains the source of live evidence.

## 3. Give the agent an evidence-first task

Name the project, time window, desired code outcome, and whether writes to
Judgment are allowed. For example:

```txt
Use the Judgment MCP tools to find checkout-agent traces with errors from the
last 24 hours. Use search_docs for current tracing guidance and load the
repo-local Judgment skill. Explain the production evidence, inspect the related
code, and propose a fix. Do not change Judgment data.
```

A grounded answer should separate three things:

1. production evidence returned by Judgment tools;
2. product guidance returned by documentation search or the skill; and
3. repository evidence from the code the agent inspected.

Use [Judgment Agent](/documentation/judgment-agent) instead when the
investigation begins in the product and page context or selected traces matter
more than local code.