Agent Toolkit

Why the Agent Toolkit exists

AI agents are powerful, but without the right context they'll confidently work with stale tables, ship changes that break downstream dashboards, or miss the alert that explains the bug you're chasing. Monte Carlo already has that context — freshness, lineage, active alerts, monitor coverage, query history, cost signals.

The Agent Toolkit is the bridge. It's a bundle of skills, hooks, and an MCP server that teaches your coding agent when and how to reach for Monte Carlo, so the observability data you already have shows up at the moment it's useful. The result: fewer broken merges, faster incident response, and monitoring coverage that keeps up with your data estate — all without you having to remember which Monte Carlo feature to call.

How the toolkit works

The toolkit is delivered as a single mc-agent-toolkit plugin that bundles everything — install once, get everything, and new capabilities arrive on the next update. Under the hood there are three building blocks that work together:

• An MCP server that gives the agent live, authenticated access to your Monte Carlo account (alerts, lineage, monitors, query history, and more).
• Skills — focused instructions that teach the agent how to do specific Monte Carlo jobs well.
• Hooks — event-driven triggers that activate skills at the right moment (for example, before you edit a dbt model, or before a commit).

Quick start

Prerequisites

• A Monte Carlo account with Editor role or above
• A supported AI coding agent (see Supported agents below)

Install the mc-agent-toolkit plugin

Claude Code (recommended — best support):

1. Add the marketplace:

/plugin marketplace add monte-carlo-data/mc-agent-toolkit

2. Install the plugin:

/plugin install mc-agent-toolkit@mc-marketplace

3. Authenticate — run /mcp, select the Monte Carlo server, and complete the OAuth flow in your browser.

4. Verify — ask Claude: "Test my Monte Carlo connection."

5. Enable auto-updates — run /plugin, select the monte-carlo-data/mc-agent-toolkit marketplace, and choose Enable auto-update. This keeps the toolkit current as new skills ship.

Cursor:

bash <(curl -fsSL https://raw.githubusercontent.com/monte-carlo-data/mc-agent-toolkit/main/plugins/cursor/scripts/install.sh)

Then restart Cursor. The Monte Carlo MCP server will prompt for OAuth on first use.

GitHub Copilot CLI:

git clone https://github.com/monte-carlo-data/mc-agent-toolkit.git

# Install hooks into your dbt project
./mc-agent-toolkit/plugins/copilot/scripts/install.sh /path/to/your/dbt-project

# Install the plugin (skills + MCP)
copilot plugin install ./mc-agent-toolkit/plugins/copilot

For other agents (OpenCode, Codex), see the plugins page on GitHub.

📘

Keeping the toolkit up to date

Third-party plugin marketplaces in Claude Code don't auto-update by default. After installing, we recommend enabling auto-updates: In Claude Code: /plugin → select the monte-carlo-data/mc-agent-toolkit marketplace → Enable auto-update

For Cursor and other agents, re-run the install script periodically or watch the GitHub repo for releases.

Orchestrated workflows

For common, multi-step situations — responding to an incident, deciding what to monitor, or sorting through an ambiguous "something feels off with my data" — the agent proactively recommends and chains the right skills in the right order. You don't have to know which individual skill to call; the workflow handles the sequencing.

You can also launch a workflow explicitly with a slash command:

Workflows are smart proactive ways to sequence the individual skills below.

Individual skills

Every capability is also available on its own. Ask your agent a direct question ("check the health of the orders table", "tune this monitor to reduce noise", "what's costing us the most in storage?") and the matching skill activates. You can mix and match — the workflows are a convenience, not a cage.

What the skills do

Skills are grouped by the job they help you do. All skills ship with the plugin and are available in every supported editor.

Trust — is this data safe to use?

Foundational pre-query and pre-build checks so the agent doesn't reach for data that isn't ready.

Incident Response — something is broken, fix it

Reactive skills for when alerts are firing, a table looks stale, or a pipeline failed.

Monitoring — plan coverage, create monitors, reduce noise

Skills for reasoning about monitor coverage and keeping monitors healthy.

Prevent — catch issues before they ship

Silent, hook-driven skills that shape code changes before and after they happen. Auto-activates on dbt/SQL edits.

See How MC Prevent works below for the full flow.

Optimize — cost and performance

Skills for cleaning up waste and diagnosing slowness in existing data assets.

Setup — onboarding and integrations

Admin-oriented skills for getting data into Monte Carlo and configuring connections.

Full details and source are in the agent toolkit repo.

Using skills standalone (without the plugin)

If your editor isn't listed above or you prefer to install skills individually:

1. Configure the Monte Carlo MCP server:

claude mcp add --transport http monte-carlo-mcp https://mcp.getmontecarlo.com/mcp

2. Install a skill:

npx skills add monte-carlo-data/mc-agent-toolkit --skill prevent

3. Authenticate via /mcp and complete the OAuth flow.

See the MCP server docs for advanced configuration options.

Supported agents

The mc-agent-toolkit plugin is available for multiple AI coding agents:

Editor compatibility: VS Code users can use Copilot CLI or Claude Code. JetBrains users can use Copilot CLI. Cursor is a standalone editor. Claude Code, Copilot CLI, and OpenCode also run in any terminal.

How MC Prevent works

MC Prevent is the flagship feature of the toolkit. When installed via the plugin, it uses hooks to enforce a safety-first workflow:

1. You edit a dbt model — the pre-edit hook blocks until a change impact assessment runs. The agent surfaces downstream blast radius, active alerts, monitor coverage, and a risk-tiered recommendation.
2. You confirm and edit — the post-edit hook records the change. The skill offers to generate monitors-as-code for new logic.
3. You commit — the pre-commit hook checks for unresolved monitor coverage gaps.
4. You validate — run /mc-validate or ask the agent to generate validation queries. Targeted SQL checks are saved to validation/<table>_<timestamp>.sql.

Without the plugin, the skill still works — it just relies on the agent following instructions rather than hooks enforcing the workflow.

How Monitor Creation works

Asking an AI agent to "add a monitor" without guardrails usually produces monitors that reference nonexistent tables, invalid fields, or the wrong monitor category for the problem. The monitoring-advisor skill prevents this by grounding every step in your live Monte Carlo workspace:

1. Category selection — the agent matches the user's intent to the right Monte Carlo monitor category (table, metric, validation, comparison, custom SQL, or one of the agent monitor types) rather than defaulting to whatever it saw most in training.
2. Table validation — the agent confirms the target table is ingested into Monte Carlo before proposing anything, using the MCP server to check in real time.
3. Field verification — for metric and validation monitors, the agent checks that referenced fields exist and that their types are compatible with the chosen category.
4. Monitors-as-code output — the agent emits monitors-as-code YAML, so the monitor is version-controlled and deployable through your existing CI/CD instead of clicked into the UI.

If you want to start from coverage instead of a specific table, run /mc-proactive-monitoring — the workflow analyzes gaps across your warehouses and use cases first, then hands off to monitoring-advisor to create the recommended monitors.

Resources

• GitHub repository — source code, detailed READMEs, and architecture docs
• Contributing — guidelines for adding skills and plugins
• MCP Server docs — Monte Carlo MCP server configuration and auth options