Skip to main content
Unlisted page
This page is unlisted. Search engines will not index it, and only users having a direct link can access it.

Overview

Use AI in Foxglove to work with your robotics data in natural language. Ask questions about the data you have open, build and refine layouts, write user scripts, and search your recordings — by describing what you want instead of doing it by hand.

Ways to use it​

You can use AI through three surfaces:

  • AI Chat — a built-in assistant in the Foxglove app. Best when you're already in Foxglove looking at a recording or layout and want fast, grounded back-and-forth.
  • MCP server — the desktop app runs a local Model Context Protocol (MCP) server, so external agents like Claude Code, Cursor, and Claude Desktop can control Foxglove. Best when you already work in another AI tool, or when an investigation spans Foxglove and other systems like your code.
  • Slack bot — the Foxglove bot in Slack. Best for quick questions about Foxglove and your organization's data — devices, Projects, sites, and events — without leaving Slack.

AI Chat and the MCP server work through effectively the same set of tools. They differ in one way that matters: AI Chat asks for your approval before an action changes anything, while an external agent's actions run as soon as it calls them.

What it can do​

The assistant works by calling tools — the same operations you'd perform in the app:

  • Understand your data — list topics, read schemas and messages, and summarize what's in the data source you have open. It can also capture an image of the app, the layout, or a single panel to see exactly what you're seeing.
  • Build and edit visualization layouts — create a layout (Foxglove's term for a saved arrangement of visualization panels) for a task, add or reconfigure panels, and validate a layout against your data.
  • Write user scripts — create and update user scripts to derive new topics from existing ones — for example, converting a custom schema into one of Foxglove's well-known schemas so it renders in the 3D or Map panel.
  • Add and edit events — create, update, and delete events to annotate moments in your data.
  • Search your data — find recordings, sessions, and events across your organization's uploaded data.
  • Navigate and play back — jump to a view, seek the playhead, and control playback.
  • Read and update devices — look up devices in your organization, and create or update them.
  • Answer product questions — search the Foxglove documentation and cite its sources.

Beyond individual tools, Foxglove ships built-in skills — task playbooks covering layout creation, panel configuration, FoxQL, user scripts, and diagnosing issues in MCAP data. AI Chat loads one when the job calls for it; you don't invoke skills directly, they shape how the assistant approaches bigger tasks. An external agent reaches the same skills, listing and fetching them as it decides they're relevant.

Example workflows​

The capabilities above combine into workflows. Some prompts to try:

  • "Build a layout to review the camera feeds and joint states in this recording." — designs a layout against your actual topics. In AI Chat, you can review the proposed layout as a diff before it applies.
  • "Why is there a gap in /imu/data around the 30 second mark? Add an event marking it once you find it." — investigates the data, then annotates what it found.
  • "Write a user script that converts /detections to a SceneUpdate so I can see them in the 3D panel." — bridges a custom schema to a panel-ready one.
  • "Find last week's recordings from this device and open the one with the most events." — searches your organization's data and navigates the app there.

Staying in control​

In AI Chat, the assistant asks before it changes anything. Read-only operations, like listing topics or reading messages, run automatically. Actions that modify state — saving a layout, editing a user script, updating an event — pause and ask for your approval first. You decide whether to allow an action once, for the rest of the chat, or always.

Every tool call also appears in the conversation, so you can expand it and see exactly what the assistant did and what it got back. See AI Chat for details.

External agents connected over MCP work differently: Foxglove runs whatever tool the agent calls, including the ones that write, without prompting you first. That covers your layouts, user scripts, and events — and devices, which reach beyond your own session, since an agent can create and rename devices in your organization. Control there comes from which agents you hand the token to, and from your agent's own approval settings. See Actions run without prompting.