Loading ToolBench
Automated Multimodal Design Reviewer & Linter for KiCad Projects. Exposes tools, resources, and prompts for autonomous AI design agents to execute deterministic DRC, heuristics, multimodal vision audits, and closed-loop repairs.
Four tools with complete input schemas and descriptions, but descriptions are verbose (200-300+ chars) and lack conciseness. All tools are READ_ONLY and well-scoped. Parameter descriptions are present but some lack constraints (e.g., vision_model accepts free-form strings instead of enum). Output schemas are partially documented in code but not formally declared in tool definitions. Error handling is present in implementation but not surfaced in tool descriptions. No tool annotations (readOnlyHint, etc.) visible in registration.
Rapidly verify that every IC power pin has a high-frequency bypass/decoupling capacitor placed within the maximum allowable distance to prevent parasitic trace inductance. Pass live=true instead of a path to audit the board open in KiCad, unsaved edits included: use it after editing through KiCad's API, e.g. with Konnect, since the file on disk only changes on save.
Retrieve a prioritized list of machine-executable ActionableFix items with precise coordinates (X, Y), layer, and target component/net for autonomous AI closed-loop layout repair. Pass live=true instead of a path to audit the board open in KiCad, unsaved edits included: use it after editing through KiCad's API, e.g. with Konnect, since the file on disk only changes on save.
Run a comprehensive 3-layer audit (deterministic DRC/ERC, spatial heuristics, multimodal vision AI) on a KiCad project. Returns structured findings, layout health score (0-100), and actionable repair proposals. Pass live=true instead of a path to audit the board open in KiCad, unsaved edits included: use it after editing through KiCad's API, e.g. with Konnect, since the file on disk only changes on save.
Execute native KiCad Design Rule Check (DRC) and Electrical Rule Check (ERC) via kicad-cli and return parsed violations with coordinates. Pass live=true instead of a path to audit the board open in KiCad, unsaved edits included: use it after editing through KiCad's API, e.g. with Konnect, since the file on disk only changes on save.
vision_model parameter accepts free-form strings instead of enum. Should constrain to: gemini-3.8-flash, fable-5, gpt-6-astra, mock. LLMs will hallucinate invalid model names.
Tool descriptions are 200-300+ characters, exceeding the 194-char baseline for A+ tools. Descriptions should be tightened to 50-150 chars with key info upfront: what it does, when to use it, what it returns.
fail_on parameter in inspect_project lacks validation guidance. Description says 'CRITICAL, WARNING, or SUGGESTION' but does not state what happens if an invalid value is passed. Should include: 'Must be one of: CRITICAL, WARNING, SUGGESTION. Invalid values will return an error.'
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | <=2025-11-25 | v2 |
Output schema not formally declared in tool definitions. Code shows _serialize_finding() and AuditResult model, but tool registration does not expose the return schema. LLMs cannot plan downstream operations without knowing what fields to expect.
No tool annotations (readOnlyHint, idempotentHint) visible in registration. All four tools are READ_ONLY and idempotent, but this is not declared in the tool definition. Agents cannot infer safety properties.