MCP server for Rigol DS1000Z / MSO1000Z oscilloscopes over raw SCPI (no VISA/UltraSigma).
Server has 11 tools with mostly complete schemas and descriptions. Naming follows verb_noun convention (get_, set_, run_, plot_). Descriptions are present but vary in quality (10-150 chars, below the 50-200 char LLM-optimized baseline). Parameter schemas are well-typed with descriptions. Key gaps: no enum constraints on string parameters (channel, mode, action, coupling, slope, sweep, depth), no output schema documentation, no error recovery guidance, and the raw 'scpi' tool exposes dangerous IRREVERSIBLE operations without confirmation or dry-run support.
live display as a PNG (Claude can SEE the scope)
measurements (Vpp, freq, duty, rise/fall, ...)
capture a trace (NORM screen or RAW deep memory) -> stats + CSV
capture + render a PNG plot (needs matplotlib)
run / stop / single / force / autoscale
identity + current settings snapshot
raw SCPI escape hatch
No enum constraints on string parameters. 'action' (RUN|STOP|SINGLE|FORCE|AUTOSCALE), 'mode' (NORM|RAW|MAX), 'coupling' (AC|DC|GND), 'slope' (RISE|FALL), 'sweep' (AUTO|NORM|SINGLE), 'depth' (AUTO|12K|120K|1.2M|12M|24M) accept free-form strings, inviting hallucinated invalid values. LLMs cannot self-correct without explicit enum constraints.
Raw 'scpi' tool marked IRREVERSIBLE with no confirmation, dry-run, or recovery guidance. Agents can send arbitrary SCPI commands (e.g. *RST to factory reset) with no safeguard. Description is minimal (20 chars). Violates confirmation-request and error-classification patterns.
No output schema documentation. Tools return data (PNG, CSV, stats, settings) but LLMs cannot plan downstream operations without knowing field names, types, and structure. Violates response-shaper pattern.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
memory depth
per-channel scale/offset/coupling/display/probe
timebase scale + offset
edge source/level/slope/sweep
Descriptions are too short (10-75 chars, below 50-200 char LLM-optimized baseline). 'measurements (Vpp, freq, duty, rise/fall, ...)' and 'timebase scale + offset' lack context for when/why to call. Agents cannot distinguish similar tools (e.g. get_stats vs get_waveform) without richer descriptions.
No error recovery guidance. Tools can fail (socket timeout, invalid channel, measurement unavailable) but descriptions do not explain what to do next. Agents have no guidance on retry, fallback, or user escalation.