This guide explains how to make an external research provider callable inside an AI harness and verify that it returns useful results before installing the multi-provider-research routing skill.
The skill is a routing layer. It tells an agent which provider to use and how to preserve evidence, but it does not create provider access.
Last substantively reviewed: September 2026. Provider endpoints, authentication methods, and harness configuration formats can change.
Public: Choosing External Research Tools for AI Agents
The skill works with any available subset of supported providers, including one. Connect a provider that fits your first task, verify it, and add others only when a distinct need arises.
Use providers that are callable, permitted, and relevant. A required provider must be attempted for its assigned role; “only” or “exactly” restricts the permitted set. If a required provider is unavailable, report the gap and follow the stated fallback policy.
| Service or platform | Practical access routes | Important note |
|---|---|---|
| Exa | Official plugin where available, or the hosted remote MCP server | The base MCP endpoint exposes Search and Fetch by default; Advanced Search and Agent are opt-in tools |
| Octen | Official Octen skills, the hosted MCP endpoint, or the local octen-mcp server |
Choose a supported interface and authenticate it. Inspect its live options and run a small read; availability and parameter shapes can differ |
| Perplexity | Hosted remote MCP server, or its optional local MCP package | Perplexity recommends the remote server when the client supports remote MCP |
| Parallel | Separate hosted Search MCP and Task MCP servers | Search/Fetch and asynchronous provider-side tasks have different schemas and lifecycle rules |
| Firecrawl | Official hosted or local MCP server | Inspect the live surface: it may expose search, scraping, mapping, crawling, parsing, agent, browser, developer, research, and monitoring tools |
| TinyFish | Curated plugin where available, or its hosted remote MCP server | The currently inspected plugin exposes Search, Fetch Content, and Web Automation; other surfaces may differ |
| X | xurl with its skill, or an authenticated X API MCP integration | These are alternative X-native interfaces; ordinary web evidence may corroborate, but it is not X-native evidence |
A remote MCP server normally requires a server URL and authentication. The provider operates and updates the server, so the reader does not need to keep a local process running. It is often the simplest portable starting point.
Use a local MCP server when:
Use an official provider skill when it contains instructions or scripts that directly call the provider API. A skill is not automatically an MCP server, and an instruction-only skill does not create connectivity.