Perplexity

API key authentication via Authorization: Bearer header. Keys are generated in the Perplexity API settings. No fine-grained scoping. Pay-per-request pricing based on model and token usage. No free tier for API access (separate from the free consumer product). No OAuth or delegated access. The auth model mirrors OpenAI's simplicity. Usage tracking is available in the dashboard. No IP restrictions on keys. For agents, the only consideration is cost management: Sonar Pro queries cost more but return higher-quality answers with more search results. Agents should select the appropriate model tier based on accuracy requirements.

Perplexity's API provides a unique capability: LLM-generated answers grounded in real-time web search results with citations. For agents, this solves the 'knowledge cutoff' problem — answers reflect current information from the web, not just training data. The Sonar model family (sonar-small, sonar, sonar-pro) offers different quality/speed/cost tradeoffs. Each response includes citations linking to source URLs. The API uses OpenAI-compatible chat completions format with system/user/assistant message structure. Search domain filtering enables constraining results to specific websites. For agents needing factual, current information — market research, news analysis, competitive intelligence, technical documentation queries — Perplexity's search-grounded approach is more reliable than standard LLMs. The limitation: the API only supports search-augmented generation, not general-purpose chat, code generation, or tool use.

Single endpoint: POST to /chat/completions at api.perplexity.ai. Request format follows OpenAI's chat completions schema: model, messages array, optional parameters (temperature, max_tokens, top_p). Model selection: sonar-small-online (fastest), sonar-medium-online (balanced), sonar-large-online (highest quality). Responses include the standard choices array plus a citations array containing source URLs. The search_domain_filter parameter limits web search to specific domains. The search_recency_filter parameter constrains results by freshness (day, week, month). No streaming support for some models. No function calling, tool use, or image input. The API surface is intentionally narrow — one endpoint doing one thing well. For agents, this simplicity means integration takes minutes, not hours.

API errors follow OpenAI's format with error.type and error.message. Rate limits are per-minute and per-day based on account tier. 429 responses include Retry-After. The main quality consideration is search result relevance: Perplexity's web search may not always surface the most relevant sources for niche or highly technical queries. Citation URLs should be verified — they reference sources but the LLM's synthesis may interpret them differently than a human reader would. No search results for very recent events (minutes-old) — the search index has a short delay. For agents, implementing citation verification (checking that cited URLs actually support the claim) adds a valuable quality layer. Timeout behavior: complex queries with many search results may take longer to respond.

Documentation at docs.perplexity.ai is minimal but sufficient — the API surface is small enough that comprehensive docs fit in a few pages. The chat completions endpoint is documented with parameters, response schema, and model list. Code examples in Python (using OpenAI client library with custom base URL) demonstrate the zero-friction integration path. The documentation explicitly states what the API doesn't support (function calling, vision, embeddings), which prevents wasted integration effort. No SDKs beyond the OpenAI-compatible client approach. Community resources are limited. The documentation's main gap: search quality tuning guidance (when to use domain filters, how recency filters affect results) is thin. For agents, the OpenAI client library compatibility means the getting-started guide is essentially: change your base_url and model name.