Loading ToolBench
An autonomous OSINT and neural intelligence reconnaissance system that leverages OSINT scanning (Maigret, Holehe) combined with Swarm deep research to extract digital footprints and background intelligence on targets. Exposes tools via Model Context Protocol (MCP) for integration with AI agents.
Single tool 'invoke_spectre_osint' has a strong, detailed description (180+ chars) explaining WHAT it does (OSINT reconnaissance), WHEN to use it (autonomous missions), and WHAT it returns (disambiguated JSON). Input schema is present with three parameters (target_name, target_handle, target_email), all with type definitions and descriptions. However, schema lacks critical constraints: no enums, no min/max lengths, no format specifications for email validation. Parameter descriptions are adequate but generic (e.g., 'optional' without guidance on when to provide each). Output schema is completely undocumented, no field listing, no structure definition. Error handling and recovery guidance absent. Tool name 'invoke_spectre_osint' is action-verb-based but somewhat opaque; 'run_osint_reconnaissance' or 'search_osint_profiles' would be clearer. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite READ_ONLY risk classification being present in metadata.
Run an autonomous SPECTRE reconnaissance mission right from your agent's context. It leverages OSINT scanning (Maigret, Holehe) combined with Swarm deep research to extract digital footprints and deep background intelligence on a target. It returns a fully disambiguated JSON payload defining the primary target persona vs overlapping personas.
Output schema completely undocumented. Tool returns 'fully disambiguated JSON payload' but no field definitions, types, or structure provided. LLMs cannot plan downstream operations or extract specific data.
Input parameters lack format constraints. 'target_email' has no email validation pattern; 'target_name' has no length bounds. LLMs will pass invalid values (e.g., 'john' as email, 500-char names) causing silent failures.
No error handling or recovery guidance. Tool description does not explain failure modes (invalid target, API rate limits, no results found) or what the LLM should do next.
Tool name 'invoke_spectre_osint' is vague. Does it search profiles, run scans, or aggregate data? Clearer names like 'search_osint_profiles' or 'run_reconnaissance_scan' would help LLMs select the tool confidently.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | <=2025-11-25 | v2 |
Tool annotations missing. Risk classification 'READ_ONLY' is present in metadata but not exposed as tool annotation (readOnlyHint). LLMs cannot infer safety properties from the tool definition alone.