Local MCP access to Firestorm's live LEAP APIs and asset inspection workflows
Firestorm MCP has 16 tools with complete schemas and descriptions. Naming is verb-forward and clear (capture_snapshot, ui_invoke_menu, viewer_call). Descriptions are detailed (avg ~150 chars) and explain WHAT, WHEN, and side effects. However, parameter descriptions are sparse or missing in several tools (viewer_call, viewer_event_poll lack detail on argument structure). Output schemas are not documented, callers cannot predict response structure. Error handling is minimal; tools state 'Dispatched does not prove effect' but offer no recovery guidance. Tool composition is sound (single responsibility), but some tools like viewer_call are overly generic gateways that bypass semantic validation.
Inspect local asset files (DAE, glTF, GLB, PNG, JPG, etc.) and extract metadata without simulator verification. Supports COLLADA geometry analysis, glTF material inspection, and image properties.
Discover all LEAP APIs and operations exposed by the connected Firestorm viewer. Refresh before using viewer_* tools after viewer updates or feature changes.
List all viewer snapshots captured by this MCP session, with metadata and file paths. Does not verify image content.
Request a viewer snapshot at specified dimensions and layer (COLOR or DEPTH). Saves to the captures directory with timestamp and UUID. Returns file metadata, actual dimensions, and quality assessment.
Report viewer connection state, API discovery status, and bridge diagnostics
Pixel-by-pixel comparison of two images. Produces a difference image and mean absolute error. Requires images to have identical dimensions. Does not verify semantic match.
viewer_call is a generic gateway that bypasses semantic validation. It accepts arbitrary API/operation names and arguments without schema constraints, inviting hallucinated operations and malformed payloads. LLMs cannot infer valid operations from the description alone.
Output schemas are not documented for any tool. Callers cannot predict response structure, forcing LLMs to guess at field names and types. This breaks tool chaining and wastes tokens on exploratory calls.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Invoke a validated menu callback by name and optional parameter. Menu names and functions are discovered via ui_list_menus. Dispatched does not prove effect.
Parse the viewer's menu_viewer.xml and list all menu entries with their click handlers and parameters. Does not invoke any menu actions.
Query a UI control's info and current value by XUI path. Returns control metadata and value if the control is value-bearing.
Read multiple UI control values from a single XUI panel file by name. Parses the panel definition and queries current values for specified controls.
Set multiple UI control values in a single XUI panel by name. Validates control existence before writing.
Call any discovered LEAP operation by API and operation name. Supply viewer fields inside arguments; transport reply/reqid are supplied automatically. Dispatched does not prove effect.
Acquire exclusive control of the viewer for a specified duration (30-1800 seconds). Returns lease expiration time. Only one client may hold control at a time.
Release exclusive viewer control. Only the controlling client may release.
Query current viewer control state: who holds the lease, when it expires, and bridge diagnostics.
Poll unsolicited viewer events (notifications, state changes) received since the last poll. Returns events after a cursor position with optional filtering.
Error handling is minimal. Tools state 'Dispatched does not prove effect' but provide no recovery guidance. When a viewer operation fails, the LLM has no actionable next step, no retry logic, no alternative tools, no diagnostic hints.
viewer_event_poll parameter 'after' lacks description of its semantics (is it a cursor, timestamp, or index?). Parameter 'pump' is optional but its purpose and valid values are not documented.
capture_snapshot parameters show_ui and show_hud lack explanation of their effect on the snapshot output. Parameter 'layer' accepts COLOR or DEPTH but does not explain what DEPTH layer represents or when to use it.