Loading ToolBench
MCP Server for Gemini Antigravity Agent - code execution, web browsing, filesystem, function calling, GitHub clone-branch-PR workflow, cron triggers, managed agents, and environment/hooks management.
Server has 22 tools with mostly complete schemas and descriptions. Naming follows verb_noun convention well (create_*, list_*, get_*, delete_*, update_*, cancel_*). Descriptions are present and substantive (100-300 chars typical), exceeding the 10-char minimum. However, several critical gaps: (1) parameter descriptions are often generic or missing detail about constraints/formats; (2) output schemas are not documented, callers cannot see what fields to expect; (3) error handling lacks recovery guidance; (4) no tool annotations (readOnlyHint, destructiveHint) despite clear risk levels; (5) some parameters like 'environment' accept 'any' type with complex nested structures that need stricter schema definition. The 'create_interaction' tool exemplifies this: prompt is clear, but 'environment' parameter accepts string OR dict OR object with no formal schema, forcing LLMs to guess structure. Tools like 'delete_interaction' and 'delete_trigger' are destructive but lack confirmation patterns or dry-run support.
Cancel an in-progress (e.g. background) interaction.
Continue an existing interaction turn with new prompt, images, or function results.
Create a managed Antigravity agent.
Create a sandbox environment for agent execution.
Create a new Antigravity agent interaction.
Create a cron-scheduled trigger for recurring Antigravity agent interactions.
Output schemas not documented. Tools return results but callers cannot see field structure, types, or what data to expect. This forces LLMs to guess and breaks tool chaining when downstream tools need specific fields (e.g., interaction_id from create_interaction).
Destructive tools (delete_interaction, delete_trigger, delete_agent, delete_environment) lack confirmation or dry-run patterns. Agents can permanently delete resources without a safety gate.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | <=2025-11-25 | v2 |
Delete a managed agent.
Delete a sandbox environment.
Delete a stored interaction.
Delete a trigger.
Fetch details of an agent by ID.
Fetch details of an environment by ID.
Fetch details of an interaction by ID.
Poll the status of a background interaction.
Fetch details of a trigger by ID.
List all managed agents.
List all sandbox environments.
List cached interactions from the session store, most recent first.
List all cron-scheduled triggers.
Submit a local function execution result back to a pending interaction (requires_action status).
Update an agent's configuration.
Update a trigger's schedule or interaction config.
Parameter 'environment' in create_interaction accepts type 'any' with complex nested structure (EnvironmentConfig dict with type, environment_id, sources, network). No formal JSON Schema constraint. LLMs cannot validate structure before calling.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. Tools marked DESTRUCTIVE (delete_*) or READ_ONLY (list_*, get_*) should declare this via annotations for agent safety.
Error handling lacks recovery guidance. No indication of which errors are retryable, which require user input, or which are fatal. LLMs cannot self-correct on failures.