Perplexity Agent API
#3 v Frameworky pro agentyPerplexity · 5× · naposledy 19. 7. 2026
The Perplexity Agent API is a managed runtime for agentic workflows released by Perplexity in March 2026, bundling web search, tool execution, and multi-model orchestration through a single endpoint. It is model-agnostic and provides access to frontier models from OpenAI, Anthropic, Google, xAI, NVIDIA, and other providers at direct provider pricing with no markup. Developers can use predefined presets (including fast search or Deep Research) or integrate custom configurations with Custom Functions/Skills to connect agents to their own backends, databases, and APIs. Recently, the API added support for "Skills"—reusable, composable capability modules that combine built-in skills (such as Office/PDF) with custom-developed Skills.
Vlastnosti
| Autonomy Level | Fully autonomous agent loop: model plans, selects tools, executes, observes results, evaluates goal completion, and iterates on its own |
| Channels | REST API/HTTP endpoint, usable via official Python and TypeScript SDKs as well as OpenAI-compatible SDKs |
| Modalities | Text input/output with tool calling (web search, URL fetch, custom functions); Sandbox API for code execution planned as an additional tool |
| Platform | Cloud API, endpoint POST https://api.perplexity.ai/v1/agent (alias: /v1/responses for OpenAI SDK compatibility); official Python and TypeScript SDKs |
| Price | Pay-as-you-go, no base fee; tool invocations: web_search $0.005/call, fetch_url $0.0005/call; model tokens at direct provider rates with no markup |
| Release Date | March 11, 2026 (official announcement/GA) |
| Tool/MCP Integrations | Built-in tools web_search and fetch_url, plus custom functions/skills to connect own backends, databases, and APIs; model fallback chains for high availability; separate MCP server offering still available |