Reading papers is
no longer a human-only job.
Elicit, the research AI behind millions of literature reviews, has moved its API and MCP server out of beta and into general availability. A new Systematic Review API now lets an agent run search, screening, extraction, and reporting end to end — straight from Claude or ChatGPT, no manual glue work required.
The "beta bridge"
is gone
Elicit's API launched in beta back in March 2026, in preview for Pro, Teams, and Enterprise users (Introducing the Elicit API). The original announcement carried the usual beta caveat — pricing, rate limits, and available features "subject to change" — and researchers or engineers had to manually copy search results and extracted data out of the Elicit app and paste them into Claude or ChatGPT to actually use them.
Just over four months later, in mid-July 2026, Elicit announced that the Elicit API and MCP server had reached general availability (The Elicit API and MCP: Powering autonomous research engines). Alongside upgraded search, it added a brand-new Systematic Review API that can run a full literature review end to end. Through the MCP server, Claude, ChatGPT, Copilot, Gemini, and any other MCP-compatible client can now reach Elicit's data directly — no human relay required.
Three APIs, one
MCP connection
Splitting the surface into three endpoints lets you hand an agent exactly the scope it needs — search only, a cited report, or the entire review pipeline.
Search API
Search across 138 million academic papers and 545,000 clinical trials, using either semantic or keyword search. An agent can hand off the "find relevant literature" step here.
Reports API
Submit a research question and get back a cited report drawing on hundreds of papers or trials, exportable as Markdown, PDF, or DOCX — ready to drop into an internal doc or a first draft.
Systematic Review API
Controls every step of a full systematic review by parameter — search strategy, screening criteria, extraction parameters, and reporting format. Built for PRISMA 2020, with every decision auditable back to a cited quote.
Accuracy, benchmarked
on 994 Cochrane reviews
Elicit published accuracy figures benchmarked against 994 Cochrane systematic reviews.
The breakdown: 95% search recall, 97% abstract screening, 99% full-text screening, 96% extraction. The Systematic Review API exposes this same accuracy over the wire — the app's benchmark numbers now apply to whatever an agent calls through the API, too. None of those figures hit 100%, though, so a second human check still belongs in any high-stakes review.
Reliable research agents need one thing above all:
access to grounded evidence itself.
Whose job actually
changes here
For engineers, integration cost collapses. Go to Claude's or ChatGPT's Settings > Connectors > Add custom connector, register the MCP server URL https://elicit.com/api/mcp, and all three APIs — Search, Reports, Systematic Review — are wired in. No more wrapper code chasing a beta spec that could change without notice. You still need a Pro-tier plan or above to call the API, and the key still comes from Elicit's developer page.
For PMs and researchers, the role shifts from reading to supervising. Design a search strategy and screening criteria up front, hand it to the agent, and it can run first-pass screening across hundreds or thousands of papers, leaving a PRISMA-auditable trail of exclusion reasons and cited quotes. Nobody has to read every paper front to back — the work becomes designing the criteria and checking the output. If you don't read papers for a living, none of this changes your day.
Get an API key
On a Pro-tier plan or above, generate a key from Elicit's developer page. Start with a light test against the Search API alone.
Register the MCP connector
Add https://elicit.com/api/mcp as a custom connector in Claude or ChatGPT and wire it into an existing agent workflow.
Run a small pilot review
Before any large-scale review, run the Systematic Review API on a topic with a known answer, and cross-check the PRISMA audit log against the actual literature.
The optimism has limits
Connection issues have already surfaced in the wild. Anthropic's own GitHub repository has an open, unresolved report from June 28, 2026 of the Elicit connector failing with "Authorization with the MCP server failed" (claude-ai-mcp Issue #503). GA status is no guarantee of a stable connection, so teams rolling this out internally should run their own reconnection tests first.
On accuracy, 96% extraction and 99% full-text screening also mean a few percent still slips through. Systematic reviews can underpin medical guidelines and regulatory decisions, so publishing or submitting agent output unchecked carries real risk. And API access still requires a paid Pro-tier plan or above — there's no free-tier way to validate this yourself. "Hand the review to an agent" more precisely means "hand the first-pass screening and drafting to an agent" — a human still needs to sign off on the final call for now.