Loading ToolBench
An AI-powered research assistant that performs iterative, deep research on any topic by combining search engines, web scraping, and Gemini large language models. Available as a Model Context Protocol (MCP) tool for seamless integration with AI agents.
Single tool with partial schema definition. Tool name 'deepResearch.run' uses dot notation (non-standard) rather than verb_object format. Input schema is present with type constraints (min/max for depth/breadth) and reasonable descriptions, but lacks comprehensive error handling guidance and output schema documentation. The tool performs a complex multi-step research operation but returns a generic MCPResearchResult without documented field contracts. Caching logic is implementation-level detail not surfaced to LLM. Risk categorization (READ_ONLY) is present but tool descriptions do not explicitly state that this is a safe, idempotent, retrieval-focused operation.
Gemini-only deep research pipeline (Google Search grounding + URL context).
Non-standard tool naming: 'deepResearch.run' uses camelCase with dot notation instead of verb-first pattern like 'run_research' or 'search_research'. LLMs expect snake_case verb_noun format to parse intent correctly.
Output schema is not documented. The tool returns MCPResearchResult with 'content', 'metadata' (learnings, visitedUrls, stats), but LLM context lacks formal schema definition of these fields. Agents cannot plan downstream operations without knowing return structure.
Error handling lacks recovery guidance. When deep research fails (error caught in try/catch), response is a bare error message string: 'Error during deep research: {message}'. Does not indicate if error is retryable, requires user input, or is fatal.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 49 | <=2025-11-25 | v2 |
| 2026-03-09 | D | 57 | - | v1 |
Tool description conflates implementation detail with interface. Description says 'Gemini-only deep research pipeline' which leaks provider coupling. LLM should not need to know Gemini is the backend; description should focus on what the tool does (performs iterative research on topics) not how.
Optional parameters 'goal' and 'flags' are defined in schema but not used in handler function signature. The async handler destructures query, depth, breadth, existingLearnings but ignores goal and flags.optionalFeatures. This creates a silent contract violation, LLM may pass these params expecting them to affect behavior, but they are discarded.
No pagination or result limiting stated. Tool description does not mention caps on learnings returned or visited URLs. If research is deep, metadata.visitedUrls could contain hundreds of URLs, bloating context. No limit or pagination parameter offered.