Hermes Adds Perplexity as a Backend for Its Web Tools
The integration leaves Hermes’s reasoning model unchanged, but its page extracts are relevant snippets—not a guaranteed complete source record.
Loading page…
The integration leaves Hermes’s reasoning model unchanged, but its page extracts are relevant snippets—not a guaranteed complete source record.
Listen to this story
Hermes now lets users separate retrieval from reasoning by routing its existing web_search and web_extract tools to Perplexity, while leaving the agent’s model provider unchanged. The integration is available from Hermes v0.21.1, uses direct HTTP rather than Perplexity’s SDK, CLI, or MCP server, and supports a separate API key. Its main limitation is evidentiary: extraction returns relevant passages, not necessarily complete pages.
Perplexity support starts in Hermes v0.21.1; v0.21.0 does not include it.
web_search is capped at 20 results and uses short context for descriptions.
web_extract may omit key details, so agents should not fill gaps from memory or inference.
Hermes Agent can now use Perplexity as the backend for its existing web search and page-extraction tools without changing the model it uses for reasoning and writing. The new option creates a clean split between retrieval and the agent model, but it comes with an evidence boundary: extraction returns relevant passages, not a complete copy of a source page.
The provider is included in Hermes v0.21.1 and later; v0.21.0 does not include it. Users can keep their current model provider for reasoning and writing while configuring a separate Perplexity API key for search and extraction.
The change affects Hermes’s web tools rather than the model running the agent. Hermes sends requests to Perplexity directly over HTTP, so the setup does not require Perplexity’s SDK, command-line interface, or an MCP server.
Hermes limits Perplexity search requests to 20 results and uses short search context for result descriptions. Extraction works differently: it returns selected material relevant to a URL, which can leave out details found on the original page.
That distinction matters when an agent is asked to make a source-grounded claim. Perplexity’s documentation gives an example in which an extract preserves a reference to “the new function” but omits the function’s name. The missing name is not established by the retrieved excerpt, so it should not be supplied from memory or guesswork.
Hermes offers a tools setup wizard as well as manual configuration. For existing installations, search-specific and extraction-specific settings take precedence over the shared web backend, so adding an API key alone may not redirect either tool to Perplexity.
web:
backend: perplexity
search_backend: perplexity
extract_backend: perplexityPerplexity recommends disabling caching and keyless fallback during testing. Otherwise, a cached response or another provider could satisfy a request after a Perplexity error; a meaningful check shows a web_search call and usable returned results.
Loading discussion...
Join the conversation
Explain when the original page should be required.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.