Google Antigravity SDK for building AI agents with tool integration, lifecycle hooks, and MCP server support
This SDK exposes 9 tools with inconsistent quality. Most tools have basic descriptions (10-50 chars) and minimal parameter documentation. Tool schemas are present but sparse: parameters lack type constraints, enums, and validation guidance. No output schemas documented. Error handling is absent, no recovery guidance or categorization. Naming is reasonable (verb_noun pattern) but descriptions are too terse to guide LLM selection. The 'ctx' parameter in record_fruit is undocumented and appears to be framework-injected, violating the principle that all parameters should be explicit and described. No evidence of pagination, batching, or composition patterns. Security: no mention of permission gates, audit trails, or secret injection.
A tool that always fails. Useful for testing error handling.
Returns a greeting for the given name.
Looks up the SKU for a given fruit.
Divides two numbers (from MCP server)
Multiplies two numbers (from MCP server)
Reads and fetches content from a URL
Records the count of fruits by SKU.
Searches the web for information
Descriptions are too terse (10-50 chars). Rubric baseline is 194 chars. LLMs cannot determine when to select tools or how they differ from similar tools.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields for chaining.
Parameter 'ctx' in record_fruit is undocumented and appears framework-injected. All parameters must be explicit and described for LLM clarity.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Views file content from the filesystem
No error handling guidance. Tools like broken_tool and read_url_content lack recovery hints, categorization (retryable vs fatal), or actionable error messages.
No pagination or result limits documented. search_web and read_url_content could return unbounded results, risking context window exhaustion.