Databricks Field Guide by TechFabric

Databricks Field Guide MCP server

Learn Databricks, practice for a certification, and answer engineering questions with the Databricks Field Guide. Ask "Help me learn Databricks" to start a diagnostic, or "Help me prepare for Data Engineer Associate" to start domain practice. The assistant uses interactive MCP Apps cards where supported and asks one question at a time in text clients. Answers include the guide's source sections and official documentation links.

https://databricks.techfabric.com/mcp

It's a remote server over Streamable HTTP: no local server to run, no sign-in, and free public learning. The optional Claude Code plugin below bundles the connection and six skills; connecting MCP alone does not install skills.

Ask how to add a Unity Catalog row filter, and the agent gets the guide's section on it, the SQL function from that section and the Databricks documentation behind it. Databricks' own docs stay the authority on syntax and limits, which is why every answer links to them.

This isn't Databricks' own MCP server. Databricks runs managed MCP servers inside a workspace that act on your data. This one is built and run by TechFabric, a Databricks partner, and it never connects to a workspace.

What's in this repository

Registry descriptors and an optional client plugin. The official MCP Registry entry names this repository, and GitHub's MCP Registry shows this README in Copilot. The server's code isn't here and isn't open source.

The live server and the registry are the source of truth. If this README and the server's tools/list ever disagree, the server is right. A weekly GitHub Action compares server.json with the registry and opens a pull request when the entry has changed.

Add it to your agent

Claude Code

To install both skills and the connection, run in Claude Code:

/plugin marketplace add Fabric-Pro/databricks-field-guide-mcp

/plugin install field-guide@techfabric-field-guide

Start a new session. The skills cover general Databricks questions, guided learning, certifications, code review, cost estimates and recent changes. They complement Databricks' development plugin, not replace workspace access or deployment tooling. Keep one Field Guide connection to avoid duplicate tools.

For just the MCP connection, without skills:

claude mcp add --transport http databricks-field-guide https://databricks.techfabric.com/mcp

Add --scope user to have it in every project.

agy

From a clone of this repository:

agy plugin install ./plugins/field-guide

Start a new session. The native agy manifest and mcp_config.json bundle the same six skills and preserve the HTTP connection without Claude-plugin conversion.

Claude Desktop and claude.ai

Open Customize, then Connectors, choose Add and then Add custom connector. Name it, paste https://databricks.techfabric.com/mcp, keep the authentication settings Claude detects, and choose Add. On a Team or Enterprise plan, an owner may have to add the connector for the organisation first.

ChatGPT

At chatgpt.com/plugins, choose + and then Add custom MCP server. Paste the URL under Connection, choose no authentication, and create it. In a chat, type @ and pick it. If Add custom MCP server isn't there, your workspace's admin has to turn it on.

Codex

codex mcp add databricks-field-guide --url https://databricks.techfabric.com/mcp

Or by hand, in ~/.codex/config.toml:

[mcp_servers.databricks-field-guide]

url = "https://databricks.techfabric.com/mcp"

Cursor

In ~/.cursor/mcp.json for every project, or .cursor/mcp.json for one:

{

"mcpServers": { "databricks-field-guide": { "url": "https://databricks.techfabric.com/mcp" } } }

VS Code

In .vscode/mcp.json, then start it from that file or the MCP view. The tools show up in Copilot's agent mode.

{

"servers": { "databricks-field-guide": { "type": "http", "url": "https://databricks.techfabric.com/mcp" } } }

GitHub Copilot

In VS Code, use the configuration above. Copilot CLI saves the server to ~/.copilot/mcp-config.json, for every project:

copilot mcp add --transport http databricks-field-guide https://databricks.techfabric.com/mcp

Copilot's coding agent on github.com takes it in the repository's Settings, under Copilot and MCP servers. Every tool is read-only, so it can have all of them:

{

"mcpServers": { "databricks-field-guide": { "type": "http", "url": "https://databricks.techfabric.com/mcp", "tools": ["*"] } } }

Anything else

Point the client at the URL with the Streamable HTTP transport and no authentication. A client that only speaks stdio can reach it through mcp-remote:

npx mcp-remote https://databricks.techfabric.com/mcp

Gemini CLI and OpenCode each have their own line in the setup guide.

Tools

All eighteen are read-only, and none calls a model: the review and the estimate are deterministic checks against the guide's text, the toolkit's generators run the extension's own code, and your agent does the reasoning. A rule applies only while the guide still says what it cites. The rules leave out what the guide doesn't argue, such as SELECT * or collect() on their own.

Three resources come with them: fieldguide://llms.txt, one line per page; fieldguide://whats-new, the last 30 days of changes as JSON; and fieldguide://page/{+path}, any published page as Markdown. Hosts that render MCP Apps show the results as an interactive view instead of plain text: since 1.3.0 that includes the review's findings and the estimate as a table.

Prompts

Hosts that support MCP prompts show these as slash commands. In Claude Code they're /mcp__databricks-field-guide__ .

Chapter and section arguments complete as you type, and so does the page resource's path. MCP completes prompt arguments, not tool arguments, so these prompts are where get_chapter's chapter names complete.

Notifications

On MCP 2026-07-28 a client can subscribe with subscriptions/listen to fieldguide://whats-new or any fieldguide://page/.... When a deploy of the guide changes a subscribed resource, such as a new margin note, a rewritten chapter or a feature moving release stage, the server sends notifications/resources/updated. Clients on 2025-11-25 have no stream to push on, so they read whats_new instead.

Every answer says when its page last changed and how long a client may cache it: 10 minutes for what's new, 15 for search and an hour for chapters and code. The index is rebuilt on every deploy of the guide.

Things to ask

• "Help me learn Databricks. Start with a short diagnostic."
• "Help me prepare for Data Engineer Associate. Start practice now."
• "Explain Unity Catalog interactively, then check my understanding."
• "How do I add a Unity Catalog row filter? Use the Databricks Field Guide."
• "What c