Loading ToolBench
Bidirectional Model Context Protocol (MCP) bridge connecting Google Gemini Spark with DeepMind Antigravity. Exposes system tools, file operations, terminal execution, Antigravity Agent orchestration, and full cross-client history/session logging to Gemini Spark AND Antigravity over MCP.
This server exposes 9 tools with basic MCP registration but suffers from significant gaps in definition quality. While tools are named with action verbs (run_, read_, write_, edit_, append_, create_, list_, send_, get_) and have descriptions present, the descriptions are often generic and lack the specificity required for LLM-driven tool selection. More critically, parameter schemas are present but descriptions within schemas are minimal or absent for many parameters, and error handling guidance is weak. The server implements OpenAgentShield security checks (a positive), but this does not offset definition deficiencies. Average per-tool score: 42.
Appends text to the end of an existing file (or creates it if it doesn't exist). Enforces OpenAgentShield sensitive file access boundaries.
Creates a full project with multiple files and optional setup commands. Demonstrates multi-file orchestration and project bootstrapping.
Performs a precise surgical search-and-replace edit in an existing file. Avoids having to rewrite the entire file when making targeted code changes. Enforces OpenAgentShield sensitive file access boundaries.
Retrieves a comprehensive report from the Antigravity Agent including task history, performance metrics, and system status. Used for observability and auditing.
Lists all Antigravity projects and conversations from the Antigravity Agent system. Returns raw conversation metadata for cross-agent awareness.
Reads the content of a file from the local filesystem with OpenAgentShield validation.
Parameter descriptions are minimal or absent. Most parameters lack detailed constraints, formats, or usage guidance. For example, 'timeout_seconds' in run_system_command has a description but does not document the min/max bounds in the description text (they are in the schema but LLMs cannot reliably parse JSON Schema constraints, they rely on text descriptions).
Output schemas are not documented. Tools return results but LLMs have no formal description of return type or structure. For example, run_system_command returns unstructured raw text with exit code; read_file returns raw file content; list_antigravity_conversations returns 'raw conversation metadata' but the schema is undefined. This forces LLMs to guess at the structure and wastes tokens parsing uncertain responses.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 51 | 2026-07-28+ | v2 |
Executes a shell/PowerShell command on the local system (e.g. python, npm, git, tests, pip). Includes safety filtering and configurable timeout (default: 180s).
Sends a task from Gemini Spark to Antigravity Agent for orchestrated execution. Enables bidirectional agent-to-agent communication and task delegation across the bridge.
Creates or overwrites a file on the local filesystem with specified content. Automatically creates parent directories if they don't exist. Enforces OpenAgentShield file safety checks and secret redaction.
Destructive/irreversible operations lack confirmation or dry-run capability. run_system_command executes arbitrary shell commands immediately (marked DESTRUCTIVE). send_spark_to_antigravity_task is marked IRREVERSIBLE but has no confirmation step. create_full_project, write_file, edit_file, and append_file all modify state without offering a preview or dry-run mode. LLMs are prone to hallucination, these tools should support confirmation before execution.
Error handling is weak. Tools do not provide actionable recovery guidance. For example, run_system_command returns '[Error] Command timed out after {timeout} seconds' but does not suggest retry strategy, increased timeout, or fallback. Missing guidance on common failure modes: file not found (read_file), permission denied (write_file), find_text not found (edit_file), non-existent project (send_spark_to_antigravity_task).
Tool composition smell: create_full_project combines file creation AND command execution. This couples two concerns and prevents agents from composing them flexibly. The tool should create files and optionally return a command list for the agent to execute via run_system_command.
Antigravity-specific tools (list_antigravity_conversations, send_spark_to_antigravity_task, get_antigravity_agent_report) expose domain jargon without explaining concepts to users. 'Antigravity', 'Spark', 'conversations' are undefined for users outside this ecosystem. Descriptions should include brief context or a link to documentation.
No pagination or result limits documented. list_antigravity_conversations and get_antigravity_agent_report use the word 'all' or 'comprehensive', implying potentially unbounded result sets. No documented page/limit parameters. Large result sets will exhaust context and increase hallucination.