RU

F6 open-sources an MCP server for its XDR: 12 tools for the SOC

Published: 2026-11-16 · Author: AI Release · @ai_release1
F6 open-sources an MCP server for its XDR: 12 tools for the SOC

⚡ The gist

🔍 What happened

The Managed XDR team at F6 has published a post on Habr and released the code of an MCP server for F6 XDR. Inside is a small Python app that runs alongside your AI client and talks to the XDR over a plain REST API. The client can be OpenCode, Claude, Cursor, Copilot in VS Code — anything that speaks Model Context Protocol.

The assistant gets twelve tools. xdr_search and xdr_count find and count records using the XDR query language. xdr_get_mapping and xdr_get_filters expose fields and available filter values. xdr_get_suggestions completes real values from a prefix — a lifesaver when the model guesses field spellings. Then come the artifacts: xdr_list_artifacts, xdr_get_artifact_structure, xdr_download_artifact, xdr_collect_artifacts, xdr_list_artifact_presets, xdr_abort_artifact_collection. Closing an alert is xdr_mark_event. The covered sections: alerts, incidents, emails, files, host events, network connections, assets, applications and versions, modules, and the audit log.

Setup is a single block in the client config, e.g. opencode.json in the project root or ~/.config/opencode/opencode.json:

```
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"xdr": {
"type": "local",
"command": ["uvx", "xdr-mcp"],
"enabled": true,
"environment": {
"XDR_BASE_URL": "https://your-xdr-host",
"XDR_API_KEY": "{env:XDR_API_KEY}"
}
}
}
}
```

🧠 Our take

F6 is upfront about it: they built the server for themselves, to take the grunt work off the SOC's first line — triaging alerts, hunting indicators, collecting artifacts, writing reports. And that, we'd argue, is the project's strongest selling point. Not "we've invented the future of security," but "we got tired of doing the same thing by hand, so here's what we actually use." The server also turned out to be useful beyond the SOC: for managers and deployment engineers who need incident details without filing a request to an analyst.

What really changes here is the barrier to entry for the XDR query language. The assistant writes the queries and figures out the event fields itself; the analyst reviews the output and makes the call. That's the most valuable part, in our view: a junior doesn't need to memorize syntax and field names, and a senior doesn't burn an evening on yet another boilerplate query. The flip side is just as obvious: a human still has to verify everything, and the cost of a mistake in a SOC is far higher than in a code editor.

One more thing worth flagging: the defaults. xdr_mark_event, xdr_collect_artifacts, and xdr_abort_artifact_collection are disabled out of the box. That's the right design, in our opinion — letting a model close alerts right away, without any burn-in period, is a bad idea. Here F6 is siding with the skeptics, and that inspires more trust than any slide deck about "an AI agent in the SOC."

Who wins? A vendor with an open MCP server gets a cheap on-ramp into other people's pipelines: the assistant already knows how to talk to their XDR. Who loses — by our hypothesis, closed platforms without this kind of plumbing: they'll have to explain why their integration costs money and a quarter of engineering work.

⚠️ What we don't know

The article promises a cost breakdown in tokens and rubles, but those numbers aren't in the portion available to us. There are no latency measurements, no accuracy figures for query generation, and no data on behavior under large telemetry volumes. Use cases 1–3 were run on a demo rig in OpenCode on the DeepSeek V4.1 Flash model in a Russian cloud — that's the company's claim, and we haven't verified it. There's a single source here: F6 itself, writing about its own product.

💬 Worth debating

  • Keep xdr_mark_event disabled, or let the assistant close alerts after a burn-in period on historical data?
  • Is an open MCP server a real hours-saver for the SOC, or just another channel into sensitive data that needs its own oversight?
  • A local model in a Russian cloud vs. a foreign API: for SOC data, what matters more — answer quality or storage jurisdiction?

Forecast: within a quarter, at least one or two other XDR/EDR vendors will ship their own MCP servers — the barrier to entry is too low to ignore.

Source: Habr / F6 company blog

🤖 AI summary
  • Date: 2026-11-16
  • Source: habr.com
  • Tags: MCP, XDR, SOC, open, cybersecurity, AI, OpenSource, neural networks
  • Key facts:
    • F6 has open-sourced xdr-mcp , an MCP server for its XDR platform: 12 tools for searching alerts , incidents, and emails.
    • Run it via uvx xdr-mcp , pip install xdr-mcp , or the Docker image ghcr.io/f6-security/xdr-mcp ; one config block in your client and you're set.
    • Artifact collection and alert closure are off by default, and file downloads only work when an XDR_DOWNLOAD_DIR folder is set.
    • The Managed XDR team at F6 has published a post on Habr and released the code of an MCP server for F6 XDR.
    • Inside is a small Python app that runs alongside your AI client and talks to the XDR over a plain REST API.
    • The client can be OpenCode, Claude, Cursor, Copilot in VS Code — anything that speaks Model Context Protocol.
    • The assistant gets twelve tools.
    • xdr_search and xdr_count find and count records using the XDR query language.
#MCP#XDR#SOC#open#кибербезопасность#AI#OpenSource#нейросети
← PreviousWhy voice input for AI never took off, even though OpenAI and Anthropic recommend itNext →Strata Squeezed a 50GB LLM Into 16GB of VRAM: 512K Context on a PC

Source: habr.com · post in Telegram