Allow LLMs to interact with Groundlight!
The server provides 8 tools with mostly complete schemas and descriptions. Tool names follow verb_noun conventions (create_detector, get_detector, list_detectors, submit_image_query, get_image_query, list_image_queries, get_image, create_alert). Descriptions are present but vary in quality and completeness. Parameter schemas are well-defined with types and constraints. However, several critical issues reduce the score: (1) Output schemas are not documented, responses rely on implicit Groundlight SDK types without clear LLM-facing schema definitions; (2) Error handling guidance is absent, tools provide no recovery hints for common failures; (3) Parameter descriptions lack format/constraint details in several cases; (4) No confirmation mechanism for destructive/state-changing operations (create_detector, create_alert, submit_image_query); (5) STDIO transport limits remote accessibility.
Create an alert for a detector that triggers actions when specific conditions are met.
Create a detector based on the specified configuration. Supports three modes: 1. Binary: Answers 'yes' or 'no' to a natural-language query about images. 2. Multiclass: Classifies images into predefined categories based on natural-language queries. 3. Counting: Counts occurrences of specified objects in images using natural-language descriptions. All detectors analyze images to answer natural-language queries and return confidence scores indicating result reliability. If confidence falls below the specified threshold, the query is escalated to human review. Detectors improve over time through continuous learning from feedback and additional examples.
Get a detector by its ID.
Get the image associated with an image query by its ID. Optionally annotate with bounding boxes on the image if available.
Get an existing image query by its ID.
No documented output schemas. Tools return Groundlight SDK objects (Detector, ImageQuery, Rule) without explicit field definitions. LLMs cannot plan downstream calls or extract required IDs without reverse-engineering the SDK.
Missing error handling guidance. No recovery hints in descriptions or error responses. E.g., if create_detector fails with invalid config, LLM has no guidance on whether to retry, adjust parameters, or ask the user.
Destructive/state-changing operations lack confirmation mechanism. create_detector, create_alert, and submit_image_query modify state but offer no dry-run or confirmation step. Agents cannot pause to verify intent before committing.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 62 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 0 | - | v1 |
List all detectors associated with the current user.
List all image queries associated with the specified detector. Note that this may return a large number of results.
Submit an image to be answered by the specified detector. The image can be provided as a file path, URL, or raw bytes. The detector will return a response with a label and confidence score.
Parameter descriptions lack format/constraint details. E.g., detector_id and image_query_id lack format guidance (UUID? alphanumeric? length limits?). 'config' parameter in create_detector lacks detailed description of field requirements.
No pagination guidance for list tools. list_detectors and list_image_queries may return large result sets. No pagination parameters (limit, offset, page_size) or guidance on result limits documented.
Missing permission/scope declarations. No documentation of what permissions/API capabilities each tool requires. No audit trail guidance or scope-based access control pattern.