Skip to content

Connect an AI Assistant

Evaluation Studio can be operated by an AI assistant. Connect a client like Claude Code and you can talk to your bench: ask it to open a board, run a capture and tell you what it sees, dig through the logs, or walk a repetitive setup sequence for you — while you watch every step happen in the application.

Three boxes: your AI client (Claude Code, Cursor, VS Code and others) talks over stdio to the evalstudio-mcp proxy installed with the app, which talks HTTP on port 28282 to Evaluation Studio and its tools; a dashed arrow labelled app_launch shows the proxy can start the app for the agent Three boxes: your AI client (Claude Code, Cursor, VS Code and others) talks over stdio to the evalstudio-mcp proxy installed with the app, which talks HTTP on port 28282 to Evaluation Studio and its tools; a dashed arrow labelled app_launch shows the proxy can start the app for the agent

The evalstudio-mcp proxy ships with Evaluation Studio and bridges your AI client to the application using MCP. It is available even when the app is closed — the agent can start Evaluation Studio itself.

Set it up

  1. Open Settings → Scripting and scroll to Connect an AI client.

  2. Pick your client — Claude Code, Claude Desktop, VS Code, Cursor, Windsurf, Visual Studio, Codex CLI, Gemini CLI, or JetBrains AI Assistant — and copy the generated snippet.

    The Connect an AI client panel: a client picker set to Claude Code, a ready-to-paste snippet with a Copy button, and a hint saying where the snippet goes

  3. Put the snippet where the panel's Where it goes line says — for Claude Code it is one terminal command; for most other clients it is a short block in their MCP configuration file.

  4. Start a conversation and ask the assistant to check the connection — it can call app_status, and if Evaluation Studio isn't running it can launch it with app_launch.

Tip

If the agent launches Evaluation Studio itself, you don't need to change any settings first — app_launch enables the API for that session automatically. Only when you want the agent to attach to a copy of the app you started do Enable HTTP API and Enable MCP Server need to be on in Settings.

What the assistant can do

The connected client sees Evaluation Studio's operations as tools, grouped by prefix:

Tools What they let the assistant do
app_* Check, start, and stop Evaluation Studio itself
ui_*, screenshot See and drive the user interface — navigate views, click controls, fill fields, read what's on screen, take screenshots
hardware_* Discover and open connected evaluation boards
analysis_* Capture data, run analyzers, export results
memmap_* Read and write device registers
plugin_* Install, update, and manage plug-ins
diagnose_* Read the application log, check performance, draft issue reports

Installed plug-ins add their own tools, prefixed with the plug-in's name — a signal-generator plug-in contributes signals_*, an instrument plug-in its SCPI console, and so on. The server also publishes task guides (as MCP resources) that teach the assistant how to drive the UI and capture data effectively — no prompting needed from you.

What it looks like

A real exchange, with an ADALM2000 attached (the tool calls the assistant made are shown inline):

You: Connect to my bench and grab a capture from the ADALM2000 — is there anything on the inputs?

Assistant: app_status — Evaluation Studio is running. hardware_open_attached_hardware — I've opened the ADALM2000 that's connected at ip:192.168.2.1. ui_navigate — now on its Oscilloscope page. analysis_capture_and_analyze — captured 8,192 samples per channel at 1 MS/s.

Assistant: Both scope channels show only a flat noise floor (roughly −0.07 to −0.03 codes, RMS 0.05) — there's no signal on the inputs. Want me to start the signal generator and capture again?

Note

Everything the assistant does is an ordinary operation: each step shows up in the application window as it happens, appears in the log, and can be captured in a recording — an agent session leaves the same trail as a person at the keyboard.

Staying in control

  • Local by default. The API only accepts connections from your own machine and has no authentication — leave Allow Remote Connections off unless you are on a trusted network and need it.
  • Scope what agents can touch. The Scripting tab's toggles are levers: turn off Enable Transaction Tools and the assistant can no longer operate the application or hardware (screenshots can be left on for a look-but-don't-touch setup).
  • Supervise hardware work. Tools that write to hardware are marked destructive as a hint to the assistant, not a lock — keep an eye on the app while an agent works on a live board, exactly as you would with a colleague at your bench.

Troubleshooting

Symptom Cause Fix
Tools error with "Evaluation Studio is not running" The app is closed; the proxy stays available and says exactly this Ask the assistant to run app_launch (or start the app yourself)
Assistant connects but tool calls fail against a running app The app was started by you with the HTTP API disabled Enable HTTP API and MCP Server in Settings → Scripting, or let the agent launch the app instead
You changed the HTTP port and the connection broke The proxy defaults to port 28282 Re-copy the snippet from Connect an AI client — it adds the right port automatically
The panel warns the proxy binary is missing Running a development build that hasn't packaged the proxy Build the full solution once, or use an installed copy of Evaluation Studio