Loading ToolBench
Structured discovery, evidence, and identity layer for AI agents. Returns FNI-ranked catalog data and evidence to reason over. Not general web search, model inference, model execution, or an inference router; never paid placement and no billing.
Free2AITools presents 5 well-structured tools with complete input schemas and detailed, LLM-optimized descriptions. All tools are verb-prefixed (search_, rank_, explain_, select_, compare_) and clearly articulate scope, constraints, and non-applicability (DO NOT USE for...). Schemas are fully specified with types, enums, required fields, and sensible defaults. However, output schemas are NOT documented, responses are described informally in text but lack formal JSON Schema definitions. Descriptions average ~250 chars and explicitly state when tools are read-only and transient-error-safe. Parameter descriptions are explicit and constraints are well-articulated (e.g., limit 1-20, enum filters). Error guidance is present (503 with Retry-After). No critical security issues detected (no credentials in params, read-only operations). Composition is clean, each tool has one responsibility and tools chain naturally via returned IDs. Main weakness: absence of documented output schemas prevents LLMs from fully reasoning about result structure.
Compare 2-25 AI catalog entities side-by-side — any catalog entity type (models, datasets, papers, tools), not models only — with FNI factor decomposition and specs where applicable. USE WHEN you already have 2+ specific entity ids. DO NOT USE to discover entities, run/execute a model, or get a recommendation; presents comparison facts for the caller to decide on, is not an inference router, never paid placement, no billing. Cold upper-range multi-paper requests may return a transient 503 (retry after the indicated delay).
Explain one entity's FNI score with the 5-factor breakdown (Semantic, Authority, Popularity, Recency, Quality); the Semantic factor is a baseline surfaced with a caveat, not a measured per-entity value. USE WHEN you already have one entity id and want its score rationale. DO NOT USE to discover entities, run a model, or get a recommendation. Read-only, no billing.
Keyword-search AI entities using the task/query text as input; returns matching catalog entries. Search results are ordered by a relevance score based on the FNI and, where term-match data is available, how well the entry matches the query. The score used for ordering may differ from the fni_score field returned in the response. The result set is bounded. Mechanically the same keyword search as free2aitools_search with the task folded in. Does NOT perform task-fit recommendation, compatibility analysis, model inference, or model execution, and is NOT an inference router. The caller makes the final selection. Read-only, never paid placement, no billing. May return a retryable transient 503; retry per Retry-After.
Output schemas are not formally documented. Descriptions state what is returned (e.g., 'returns matching catalog metadata', 'returns FNI factor decomposition') but no JSON Schema output specifications are provided. LLMs cannot reason precisely about response structure, field types, or chaining requirements.
Tool descriptions contain inline guidance ('DO NOT USE for...') but lack explicit per-tool error handling patterns or recovery guidance for common failure modes beyond the generic 503 retry note. Agents cannot distinguish between unrecoverable vs. retryable errors in all cases.
The free2aitools_select_model tool accepts nested constraints object with optional subfields (max_vram_gb, max_params_b, license, etc.) but the parameter description does not clearly state the consequences when no constraints are provided or when conflicting constraints are passed. This could lead to ambiguous LLM behavior.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | B | 74 | 2026-07-28+ | v2 |
Keyword discovery over the catalog of AI models, datasets, papers, and tools. Returns matching catalog metadata. Search results are ordered by a relevance score based on the FNI and, where term-match data is available, how well the entry matches the query. The score used for ordering may differ from the fni_score field returned in the response. The result set is bounded. The Semantic factor is a query-time baseline (fni_s null with a note), not a live per-entity measurement. USE WHEN discovering which AI entities exist for a topic. DO NOT USE for general web search, to run/execute a model, to get a generated answer, or to route to an inference provider. Read-only, never paid placement, no billing. May return a retryable transient 503 under cold-path/fallback budget limits; retry per Retry-After.
Filter the catalog by declared hardware/license metadata and return FNI-ranked candidate entries. USE WHEN you have concrete constraints (VRAM, params, license, context length, local-runnability). Constraints are metadata/heuristic filters over stored fields, NOT verified compatibility analysis, model inference, or model execution, and this is NOT an inference router. The caller is responsible for the final selection. Read-only, never paid placement, no billing.
Parameter 'id' in free2aitools_explain and 'ids' in free2aitools_compare accept arbitrary string values with advisory text ('use the returned id verbatim; do not hard-code or invent catalog ids') but no formal validation or enum specification. LLMs may still attempt to fabricate IDs if not carefully instructed.