Loading ToolBench
CLI, MCP server, and Anthropic/OpenAI API-compatible interface for Perplexity AI. Provides web search and AI model querying through multiple interfaces.
Server has 22 tools with consistent naming (verb_noun pattern: pplx_query, pplx_ask, pplx_usage) and descriptions present for all. However, descriptions vary significantly in quality, some are excellent (pplx_smart_query: 287 chars with clear intent guidance), others are minimal (pplx_usage, pplx_auth_status: 0 input params, sparse context). Input schemas are visible and typed (string, boolean defaults provided), but output schemas are entirely undocumented, LLMs cannot predict response structure. Parameter descriptions exist but lack constraint details (e.g., source_focus accepts 'aliases, raw source IDs, or comma-separated list' but no enum or format spec). Error handling and recovery guidance absent. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classification in metadata.
Quick Q&A with auto model. COSTS 1 PRO SEARCH QUERY. Prefer pplx_smart_query(intent='quick') for simple lookups (Sonar 2 first).
Complete authentication with the 6-digit code.
Send verification code to email for authentication.
Check current authentication status.
Claude Opus 4.7 — Anthropic's most advanced reasoning model. COSTS 1 PRO SEARCH QUERY. Requires Max subscription.
Claude Opus 4.7 Thinking — Anthropic's most advanced reasoning model with extended thinking. COSTS 1 PRO SEARCH QUERY. Requires Max subscription.
Output schemas completely undocumented. No tool returns a documented response structure. LLMs cannot predict what fields to extract or how to chain results to downstream tools.
Parameter constraints not formalized. source_focus accepts 'aliases, raw source IDs, or comma-separated list' but no enum or regex pattern. LLMs cannot validate input and may pass invalid values.
Tool annotations missing. Despite clear risk classification (READ_ONLY vs WRITE), no readOnlyHint or destructiveHint annotations present. Agents cannot determine safety without reading descriptions.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | <=2025-11-25 | v2 |
Claude Sonnet 4.6 — Anthropic's fast model. COSTS 1 PRO SEARCH QUERY.
Claude Sonnet 4.6 Thinking — Anthropic's fast model with extended thinking. COSTS 1 PRO SEARCH QUERY.
Deep Research — in-depth reports. COSTS 1 DEEP RESEARCH QUERY (limited monthly pool, typically 5-10 total). Only use when the user explicitly requests deep research.
Gemini 3.1 Pro Thinking — Google's most advanced model with extended thinking. COSTS 1 PRO SEARCH QUERY.
GPT-5.4 — OpenAI's versatile model. COSTS 1 PRO SEARCH QUERY.
GPT-5.4 Thinking — OpenAI's versatile model with extended thinking. COSTS 1 PRO SEARCH QUERY.
GPT-5.5 — OpenAI's latest model. COSTS 1 PRO SEARCH QUERY. Requires Max subscription.
GPT-5.5 Thinking — OpenAI's latest model with extended thinking. COSTS 1 PRO SEARCH QUERY. Requires Max subscription.
Kimi K2.6 — Moonshot's advanced model. COSTS 1 PRO SEARCH QUERY.
Kimi K2.6 Thinking — Moonshot's advanced model with extended thinking. COSTS 1 PRO SEARCH QUERY.
Nemotron 3 Super — NVIDIA's Nemotron 3 Super 120B model with extended thinking. COSTS 1 PRO SEARCH QUERY.
Query Perplexity AI with explicit model selection. COSTS 1 PRO SEARCH QUERY per call. Prefer pplx_smart_query for automatic quota-aware routing. Use this only when you need a specific model or thinking mode.
RECOMMENDED DEFAULT TOOL. Quota-aware query — checks limits and picks the best model automatically. USE THIS FOR EVERY QUERY unless the user explicitly requests a specific model. Default to intent='quick' for most lookups — it routes to Sonar 2 when appropriate. Only escalate intent when the question genuinely requires it. Response includes a metadata block showing the model used, routing reason, and current quota snapshot.
Sonar 2 — Perplexity's latest in-house model. Subject to your plan and Perplexity usage counters (see pplx_usage).
List live sources/connectors available on the account.
Check remaining rate limits and quotas for the current session.
Minimal descriptions for utility tools. pplx_usage (18 chars), pplx_auth_status (34 chars) lack context on when/why to call them. Below 20-char threshold triggers hard cap.
No error handling or recovery guidance. Descriptions mention quota costs but do not explain what happens on quota exhaustion, auth failure, or API errors. LLMs cannot self-correct.