MCP
MCP servers for research assistants in 2026
Direct answer: MCP matters because AI assistants need callable research tools, not just webpages to read. AutoSearch exposes a public MCP endpoint at https://exatoshi.uk/mcp so agents can launch research, check run status, fetch generated papers, verify DOIs, search evidence, and list prior runs through authenticated tool calls.
Why MCP changes research assistants
Research workflows are becoming agentic. A user no longer wants to copy a question into one tool, wait for a result, paste citations into another tool, and then ask a general model to rewrite the text. The assistant should be able to call the research system directly, inspect the output, and continue the conversation with evidence-aware context.
The Model Context Protocol gives AI clients a standard way to discover and call tools. For research, that means a model can ask AutoSearch to run a literature review, poll status, retrieve the paper, or verify a DOI without scraping an HTML page. This is especially important for Claude Desktop, IDE assistants, and emerging ChatGPT-style connectors that need structured tool results.
The AutoSearch endpoint
AutoSearch publishes its MCP server at /mcp and its manifest at /.well-known/mcp. The server requires an API key, passed through the x-api-key header, because research runs can contain private questions, uploaded context, and account-specific credit usage. Public documentation remains on /developers, while /api redirects there for developers who expect that path.
The current tool set is deliberately narrow: run research, get run status, fetch a generated paper, verify DOI, search evidence, and list runs. Those tools cover the lifecycle from question to manuscript without exposing administrative controls to an agent by default.
This narrow scope is a design choice. A research agent should not need shell access, database credentials, or production restart permissions to answer a literature question. It needs a stable contract: launch a bounded job, return a run identifier, expose status, and retrieve the resulting paper or evidence rows. By keeping the MCP layer focused on research actions, AutoSearch can be useful inside assistants while keeping operational controls outside the model's normal reach.
{
"mcpServers": {
"autosearch": {
"type": "http",
"url": "https://exatoshi.uk/mcp",
"headers": {
"x-api-key": "YOUR_API_KEY_HERE"
}
}
}
}Claude Desktop and ChatGPT-style setup
In Claude Desktop, the practical setup is to add the AutoSearch server to the desktop configuration file and restart the client. In ChatGPT-style agent products, the equivalent pattern is to register the remote tool endpoint and provide the authentication header in the connector configuration. The user should then see AutoSearch tools available inside the model environment.
For safety, the API key should be scoped to the user account and rotated if it is exposed. The agent should not receive production administrator credentials, database credentials, or unrelated service tokens. Research automation is useful only if the boundary between public evidence, private context, and operational control remains clear.
What a good research MCP tool returns
A good research tool should return structured data that the assistant can reason over without guessing. For a run status call, that means a run ID, status, frontier score, experiment count, summary, research question, and report URL. For a DOI verification call, it means the normalized DOI, verification status, title, venue, year, and warning fields. For a paper retrieval call, it means markdown or a stable artifact URL. These fields make the assistant more useful because it can cite the tool output directly instead of inventing connective tissue.
Why MCP improves citation quality
When an LLM reads a normal webpage, it has to infer what is stable and what is decoration. When it calls a tool, it receives structured output: run ID, status, evidence rows, DOI verification result, or paper markdown. That reduces ambiguity and makes citation of AutoSearch as a source more likely, because the assistant can quote concrete fields rather than summarize vague marketing claims.
MCP also lets AutoSearch become part of a larger research chain. A user can ask an assistant to compare tools using the comparison page, run a review through MCP, inspect methodology on /methodology, and estimate cost on /pricing. The page and the tool reinforce each other.
For 2026 GEO strategy, that is the key point. Pages make public claims discoverable; tools make those claims operational. If an assistant can both read that AutoSearch verifies DOI citations and call a DOI verification tool, the citation surface is stronger than a static feature list alone.