llm-perplexity-agent

Perplexity's Agent API — multi-step web research with citations — as a model for Simon Willison's llm CLI.

llm install llm-perplexity-agent
llm keys set perplexity        # skip if you already set this for llm-perplexity
llm -m perplexity-agent 'What changed in MCP 2026-07-28?'

Why

llm-perplexity wraps Perplexity's older Sonar chat models. This wraps the Agent API (POST /v1/agent) — the multi-step endpoint that runs its own searches, fetches pages, and synthesizes one cited answer. Different product, different endpoint.

Options

llm -m perplexity-agent -o preset xhigh -o recency week 'Latest on X'
llm -m perplexity-agent -o domains 'nasa.gov,-reddit.com' 'Artemis status'
llm -m perplexity-agent -o timeout 600 'Something genuinely deep'

Poll progress goes to stderr, so the answer pipes cleanly.

A note on prompt injection

Retrieved web content is untrusted, and this tool feeds it to a language model. The companion MCP server always wraps its answers in a delimiter marking them as untrusted data, because there the answer goes straight into a model that is holding tools — injected instructions could cause actions.

Here the answer goes to your terminal, and llm runs no tool loop by default, so that wrapping is off by default. Turn it on when the output is headed somewhere that matters:

llm -m perplexity-agent -o spotlight true 'Research X' | llm -m gpt-5 'Summarise'

It is a mitigation, not a fix. See SECURITY.md.

Sibling

This is one of two adapters over the same client. The other is perplexity-agent-mcp, an MCP server for Claude Desktop, Claude Code, VS Code and other MCP clients. Both are released together at the same version from one repository.

BSD-3-Clause. Copyright (c) 2026, Constantin Gonzalez.