How the Buron MCP server works

The Buron MCP server lets your AI editor talk to Buron directly. Instead of switching between your editor and the Buron dashboard, the agent in your editor can query your ad data, run dashboards, read and write knowledge files, and draft account changes, all from the same conversation where you're working.

It works with anything that speaks the Model Context Protocol: Claude Code, Cursor, Codex, GitHub Copilot, and claude.ai itself (add it as a custom connector under Customize, then Connectors). You connect once, sign in through the browser, and the client handles the rest.

The server URL is the same for everyone:

https://app.buron.ai/api/mcp

What your editor's agent gets

The toolset covers seven areas. The tools reference lists every tool with its exact behavior.

Datasets: the starting point. listDatasets, describeDataset, and queryDataset let the agent answer data questions without writing a query language: it picks measures, dimensions, filters, and a date range, and gets typed columns back. This is the same semantic layer that powers Explore in the app.

A live Google Ads escape hatch. When no dataset covers the question, runGoogleAdsGaql runs a read-only GAQL query against your live account: bidding strategies, conversion actions, ad approval status, the config graph as it is right now.

Saved queries and dashboards. Save a semantic query for reuse, run it later, assemble dashboards from saved queries, and run any dashboard for a date range with full panel data back.

Knowledge files. Read, write, list, search, and pattern-match the files in your team's knowledge layer.

Skills. List and read Buron's editor skills, the same set npx skills add buron-ai/skills installs.

Account-change proposals. The write rail: the agent can draft typed, guardrailed Google Ads changes as proposals. Nothing is ever applied by the agent; a human approves every proposal in the app.

Integration status. Connection state per integration, including the Google Ads account topology: which accounts are enabled and which conversion action is primary.

How it relates to the CLI

The MCP server and the CLI are two doors into the same API. The MCP server works from inside an AI conversation, with browser sign-in. The CLI works from your terminal and from CI, where there's no browser: it authenticates with a session from buron login or a BURON_TOKEN you mint with buron token create.

buron setup wires both worlds at once: it writes the MCP config for your editors and installs the skills.

Stateless by design

Every request to the MCP server is independent. There's no session to manage, no connection to keep alive, and nothing to clean up. Your editor connects, makes a tool call, gets a response, and moves on.