KiCadStamp MCP server: read the live KiCad board, apply validated placement config, and (opt-in, raw) move footprints.
KiCadStamp MCP server has 10 tools with complete input schemas and descriptions. Naming follows verb_noun convention (get_, list_, apply_, move_). Descriptions are detailed (100-300 chars) and explain WHAT, WHEN, and risk level. However, output schemas are not documented in the visible code, LLMs cannot predict response structure. Error handling is mentioned in descriptions ('Fails clearly', 'never raising') but recovery guidance is absent. No tool annotations (readOnlyHint/destructiveHint) despite clear risk stratification (READ_ONLY vs WRITE vs DESTRUCTIVE). Parameters lack enums and format constraints (e.g., ref_prefix is free-form string; layer is free-form string). The raw write tool (kicad_raw_move_footprint) has a safety check (expected_board_name) but no dry-run or confirmation pattern.
RAW (high-risk) write — bypasses the validated config layer entirely. Moves one footprint by ref to an absolute position/rotation directly over kipy, without registry/FORK-1/dependency-order protection. Only registered when KICADSTAMP_MCP_ALLOW_RAW_WRITE=1. The board-identity guard ALWAYS runs before writing: expected_board_name (REQUIRED) is the board you expect to be open, and the write is refused when a different board is open in KiCad; the connected board is always reported. Confirm carefully — the host's permission gate decides.
Apply a KiCadStamp placement config through the existing VALIDATED path — the same as the CLI 'apply' and the GUI Redraw: full pre-validation (board identity, FORK-1, 'never guess silently'), registries and dependency ordering. config_path is the .sexp/.json profile (e.g. profiles/3ch-awg-tia-v103-test/3ch-awg-tia.sexp). With dry_run=true it only plans and reports — nothing is written. Returns the dry-run report or the run's log lines; fatal validation messages are returned verbatim.
Read the identity of the board currently open in KiCad: the board name (stem of the .kicad_pcb) and the KiCad version. Read-only, validated.
Get one footprint in detail: position/rotation/layer, Role/Cluster fields, its pads (number, net, position) and the nets on its pads. Read-only, validated. Fails clearly when the ref is not on the board.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). Descriptions mention what is returned ('refs and uuids by kind', 'full detail') but JSON schema is absent from visible code.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools are manually risk-labeled in descriptions ('READ_ONLY', 'WRITE', 'DESTRUCTIVE') but MCP spec annotations are absent, preventing automated safety filtering.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Resolve a list of board item uuids to detailed records (tracks, vias, footprints). Each requested uuid appears exactly once in the result: found items get their full detail (kind, net, layer, mm coordinates); items not on the board return {'kind': None, 'found': False} — never raising and never silently skipping. Prefer this over listing the whole board when the uuids already come from kicadstamp_get_selection. Read-only, validated.
What the PCB editor currently has selected (groups expanded): refs and uuids by kind. Useful to see what was clicked in the KiCad GUI. Read-only, validated.
List footprints on the board (ref, Role/Cluster fields, position in mm, rotation in degrees, layer). Optional ref_prefix filters by reference prefix (e.g. 'U' lists every U*). Read-only, validated.
List all board net names. Read-only, validated.
List track segments with optional net and/or layer filters (e.g. net='GND', layer='F.Cu'; layer matches the string layer name). On a large board this can be large — prefer the net/layer filters, or kicadstamp_get_items_by_uuid when you already have uuids from kicadstamp_get_selection. Read-only, validated.
List vias with an optional net filter (e.g. net='GND'). On a large board this can be large — prefer the net filter, or kicadstamp_get_items_by_uuid when you already have uuids from kicadstamp_get_selection. Read-only, validated.
Parameters lack enums and format constraints. 'ref_prefix', 'layer', 'net' are free-form strings with no enum or pattern. LLMs cannot validate input and may pass invalid values (e.g., layer='InvalidLayer').
No dry-run or confirmation pattern for destructive write (kicad_raw_move_footprint). Tool has a safety check (expected_board_name) but no way to preview changes before committing. Agents cannot safely test moves.
Error handling lacks recovery guidance. Descriptions state 'Fails clearly' and 'never raising' but do not explain what the LLM should do next (retry, ask user, call alternative tool). No error classification (retryable vs user-fixable vs fatal).