This reference compares practical roles for Octen, Exa, Perplexity, Parallel, Firecrawl, and TinyFish. It also explains when authenticated X-native evidence belongs in a research workflow.
These services overlap, but they are not interchangeable. Choose according to the evidence and output the task requires—not according to how many providers are available.
Last substantively reviewed: September 2026. Provider tools, schemas, availability, and commercial terms can change. Inspect the tools exposed in the current harness before relying on a capability described here.
Public: Choosing External Research Tools for AI Agents
| If you need… | Start with… |
|---|---|
| A focused web or news lookup | Octen Search |
| Several interpretations of the original question | Octen Broad Search |
| Web images or relevant moments in videos | Octen Image Search or Video Search, when callable |
| Semantically precise web discovery | Exa Search |
| Exact domains, dates, categories, text, or subpage controls | Exa Advanced Search |
| Ranked URLs or recent-news leads without a generated answer | Perplexity Search |
| A quick, focused factual answer or concise cited explanation | Perplexity Ask |
| A comparison or decision analysis against explicit criteria using current web evidence | Perplexity Reason |
| A deep multi-source investigation | Perplexity Research |
| Answer-oriented web results with paired search and fetch sessions | Parallel Search |
| One substantial provider-generated research report | Parallel Deep Research |
| The same structured enrichment task repeated across several inputs | Parallel Task Group |
| Structured list-building or specialized provider-backed fields | Exa Agent and relevant Connect providers |
| Prediction-market odds, history, order books, or public position data | Exa Agent with Polymarket Connect, when exposed |
| Reading a page found by Octen | Reuse sufficient full content returned by Search or Broad Search; use Octen Extract when the content is missing, incomplete, or needs a fresher read |
| Reading a known or Exa-discovered page | Exa Fetch |
| Discovering a site’s relevant URL inventory before extraction | Firecrawl Map |
| Reading or structurally extracting one known webpage | Firecrawl Scrape |
| Collecting a bounded set of related pages from one site | Firecrawl Crawl |
| Parsing a supported document | Firecrawl Parse |
| Compact web, news, or research-paper discovery | TinyFish Search |
| Browser-rendered extraction or interactive website work | TinyFish page extraction or browser automation |
| Authenticated posts, users, trends, counts, or bookmarks from X | An authenticated X-native interface |
Apply provider constraints in this order: explicit inclusions or exclusions, named-provider requests, materially matched specialized data, then a generic discovery default. A named provider is required for its requested role; “only” or “exactly” makes the permitted set exclusive. Add a complementary provider only within the authorized scope.
An MCP connection lets the harness-selected model call a provider. It does not automatically make that model the one used inside every provider operation.
| Operation | Typical boundary |
|---|---|
| Search, fetch, or extraction | The provider returns ranked results or page content. The harness-selected model interprets the evidence and normally writes the final response. |
| Answer, reasoning, research, or agent run | A model or workflow inside the provider may generate prose or structured output before returning it to the harness. |
Search and extraction providers may still use machine-learning systems for ranking, rendering, extraction, or classification. The distinction is about who produces the visible analysis or narrative answer.
When the harness-selected model must perform all visible reasoning and synthesis, allow source retrieval and faithful extraction but exclude provider-generated answers. Inspect the selected output mode as well as the tool name: a scrape or fetch tool may also offer generated summaries or freeform answers. Ordinary lookups can default to retrieval and harness synthesis unless the user requests a different approach.
For changing facts, check available cache controls and retain the source’s observation or reporting date. Fetching a page today does not establish that its contents are current. Disclose material freshness uncertainty.
When provider-generated synthesis is allowed, require the final response to identify the provider tool and any model or preset exposed in the completed result.
Use Octen for focused web or news discovery, multi-angle expansion from the original question, page content returned with search, selected page extraction, or media discovery when available.
search_options; inspect the schema for other interfaces.