Loading ToolBench
Standalone MCP server for native Sentaurus batch experiments
Sentaurus MCP has well-structured tool definitions with clear naming (verb_noun pattern) and comprehensive parameter schemas. All 8 tools have descriptions and typed input schemas. However, descriptions are often terse (10-50 chars), lacking context on WHEN to use tools or dependencies between them. Output schemas are not documented. Error handling guidance is absent. Parameter descriptions exist but lack format constraints, ranges, or examples of valid values. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.
Prepare input files; each stage specifies tool and input_file. Tool commands come from operator config. Does not run. Native batch folder, not an SWB project.
Read-only current-host readiness; works without configuration. Does not test licenses or run simulations.
Generate remote setup guidance and read-only client config. Never connects, deploys or accepts passwords. VNC-only access is unsupported.
Read execution state, process IDs, timing, manifest and validation label.
Index native files; does not read large TDR files into model context.
List persistent experiments, paginated. Read-only.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields (e.g., what does get_experiment return? Does it include run_id for chaining to read_log?).
Descriptions are terse (10-50 chars) and lack context. Missing WHEN to use each tool, dependencies, and prerequisites. E.g., 'Read-only current-host readiness' doesn't explain when diagnose_environment should be called vs get_connection_guide.
Parameter descriptions lack format constraints and ranges. E.g., 'offset' and 'limit' in list_experiments have no min/max bounds; 'stage_index' in read_log has no validation guidance. LLMs may pass invalid values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Read bounded log increments; cursor offset is bytes.
Launch a prepared experiment asynchronously. Duplicate requests do not restart it. Poll status; no fixed runtime kill.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk profiles. create_experiment and submit_experiment are WRITE operations; list_* and get_* are READ_ONLY. Annotations would guide agent planning.
No error handling guidance. Tools lack recovery hints (e.g., 'If run_id not found, call list_experiments() to discover valid IDs'). Agents cannot self-correct on failures.