How to Run a Literature Review in Claude or ChatGPT with Verified Citations
On this page
- Add Valyu to Claude or ChatGPT as a custom connector using https://mcp.valyu.ai/mcp. You don't need any code.
- Ask one narrow question, and request an evidence table with study design, population, effect size, a supporting passage and a PMID or DOI.
- Check every citation against PubMed and the trial registry: does the paper exist, has it been retracted, and do the numbers match?
- Ask what the evidence rests on. In our GLP-1 test, four meta-analyses agreed, but one trial sat behind most of the result.
- This speeds up a narrative literature review. It does not replace a formal PRISMA systematic review.
If you're an agent: Set up Valyu and make your first search.
If you're a human: Create an account and get an API key.
MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.
What will the review show?
Suppose you are a medical writer drafting a background section. You need one defensible paragraph, and every reference has to survive a colleague's check. You ask:
Do GLP-1 receptor agonists reduce major adverse cardiovascular events (MACE) in adults with overweight or obesity but without diabetes?
A list of papers would not settle this. You need an evidence table that separates direct from indirect evidence, shows which sources you can check, and tells you what you still can't claim.
Step 1: connect Valyu to Claude or ChatGPT
Create a free account at platform.valyu.ai. New accounts get $10 of free credits, or $20 with a company email, and no credit card is required.
- Claude: Open Customize, then Connectors, and click Add custom connector. Name it Valyu, paste https://mcp.valyu.ai/mcp, click Add, then click Connect and sign in.
- ChatGPT: Open Plugins on the web, click +, then Add custom MCP server. Name it Valyu, paste the same address under Connection, leave authentication on OAuth, tick the confirmation box, click Create as a plugin and sign in. In a new chat, type @ and pick Valyu.
Work accounts may need an admin to add it first. The chat apps guide has the current steps for each plan.
Step 2: set the question's boundaries
Before you send anything, write down what counts as evidence. For this question:
- Population: adults with overweight or obesity and no diabetes.
- Intervention and comparison: a GLP-1 receptor agonist against placebo.
- Outcome: MACE, with the effect size and confidence interval.
- Evidence types: randomized trials first, then meta-analyses. Trials in people with diabetes count as indirect, and preprints must be labeled as not peer reviewed.
Step 3: ask for an evidence table
A review has more than one question in it, so ask for DeepResearch by name. Paste this into Claude or ChatGPT:
Use Valyu DeepResearch in fast mode to review whether GLP-1 receptor agonists reduce major adverse cardiovascular events in adults with overweight or obesity but without diabetes. Find randomized trials and meta-analyses. For each study give the design, population, intervention, follow-up, primary cardiovascular outcome with effect size and confidence interval, a short supporting passage, and the PMID or DOI. Separate direct evidence from indirect evidence such as trials in people with diabetes. Do not cite a paper unless it was retrieved. State gaps and mark anything unverified. Ask me to confirm before you start.
Then widen the sources with a second prompt. A review also needs to know what is still running:
Use Valyu to search clinical trial records and preprints on this question. Give each trial's NCT ID and status, and label every preprint as not peer reviewed.
Step 4: read the evidence
The report named one dedicated trial and four meta-analyses as direct evidence. We checked these figures in the source papers:
| Study | Design | MACE result (95% CI) | Worth knowing |
|---|---|---|---|
| Randomized trial, 17,604 adults with cardiovascular disease | HR 0.80 (0.72–0.90) | Semaglutide 2.4 mg; mean follow-up 39.8 months | |
| Meta-analysis of 24 trials, with a subgroup without diabetes | RR 0.81 (0.74–0.88) | Authors say the results were "mostly derived from one single trial" | |
| Meta-analysis of 17 trials, 34,419 participants | RR 0.80 (0.72–0.89) | Authors say secondary outcomes were "largely influenced" by SELECT | |
| Meta-analysis of 29 trials, 37,348 adults | RR 0.80 (0.72–0.89) | Without SELECT: RR 0.58 (0.34–0.99) | |
| Meta-analysis of 16 trials, 23,467 adults | RR 0.80 (0.72–0.89) | SELECT contributed 77% of the MACE weight |
The registry search added what papers alone could not. SELECT's registry record shows it is complete, with 17,604 participants enrolled. SURMOUNT-MMO, a tirzepatide outcomes trial, was listed as active and not recruiting, with primary completion estimated for October 2027. So there is more evidence still to come.
Step 5: verify every citation
A citation only tells you where to check. For each source, confirm that:
- The paper exists, and the title, journal and year match its PubMed record or DOI.
- It has not been retracted. PubMed marks retracted publications, and journals post notices.
- The effect size, interval and population match the paper. Check the subgroup too, not just the headline.
- It is the right kind of source. A review or commentary describing a trial is not the trial itself.
- Any trial ID resolves to the same study on ClinicalTrials.gov.
We ran this check on all 13 sources in the report. Each one resolved to a matching PubMed record, and none was marked as retracted on the day we checked. The figures in the table above matched the abstracts and open-access full text. The check also turned up limits to plan for:
- The main trial was read from its abstract only. The report said it could not retrieve the full SELECT paper. Read it yourself before citing subgroup or safety detail.
- The report's citations were all PubMed links. Registry records and preprints came from the separate searches, which is why the second prompt matters.
- One meta-analysis needs caution. The report flagged that the Cureus paper's study table lists Marso 2016, the LEADER trial in people with type 2 diabetes, with a sample size of 75, despite the paper's non-diabetic criteria. We confirmed that entry. The paper also reports a significant stroke reduction that Yin and Kelkar do not.
Figure 1. Why the pooled results are not independent confirmations, using figures as reported in each paper.
What can you conclude from the evidence?
| What the sources show | A careful description | What you cannot conclude |
|---|---|---|
| SELECT: HR 0.80 for semaglutide 2.4 mg | MACE fell in adults with existing cardiovascular disease, a BMI of at least 27 and no diabetes | That people without cardiovascular disease get the same benefit |
| Four meta-analyses at 0.80–0.81 | Pooled estimates agree, and SELECT carries much of the weight | That four papers independently confirm the result |
| Without SELECT: RR 0.58 (0.34–0.99) | The remaining evidence is far less precise | That the true effect is larger than SELECT found |
| Benefits in trials of people with diabetes | Indirect evidence from a different population | That the same effect applies without diabetes |
The decision: cite SELECT as the primary source, and cite the meta-analyses as summaries that lean on it. Don't write that GLP-1 drugs as a class have proven cardiovascular benefit in everyone with obesity.
The next action: read the full SELECT paper, resolve the stroke disagreement between the meta-analyses before mentioning stroke, and check SURMOUNT-MMO when it reports.
The outcome: a paragraph you can defend: "In adults with overweight or obesity, established cardiovascular disease and no diabetes, semaglutide 2.4 mg reduced MACE compared with placebo (HR 0.80, 95% CI 0.72–0.90). Evidence in people without cardiovascular disease, and for other agents, is still limited."
Which Valyu sources help with a literature review?
A review rarely fits in one database. Name the source in your prompt when it matters, such as "on PubMed" or "in clinical trial records".
- PubMed: peer-reviewed papers and identifiers.
- Clinical trials: NCT IDs, status and enrollment, including trials with no published results yet.
- medRxiv and bioRxiv: recent preprints, which have not been peer reviewed.
- DailyMed: drug labels, which appeared in our tirzepatide search.
- Wiley Health & Life Sciences: licensed full text, enabled per organization for eligible startups. See the Wiley guide for access.
How to adapt the review
Change the population, intervention and outcome, and keep the evidence rules. Outside medicine, swap trials for the strongest study design in your field and use DOIs instead of PMIDs. If you already have your papers, the selected-paper comparison guide shows how to request a table and CSV. To build this into a product, see building a biomedical research agent.
FAQ
Can Claude or ChatGPT replace a systematic review?
No. A PRISMA systematic review needs a registered protocol, reproducible database searches, documented screening and usually two independent reviewers. This workflow suits narrative reviews, background sections and scoping a question before formal work begins.
How do I check whether a citation is fake?
Search the PMID or DOI directly, then compare the title, authors, journal and year. If the identifier doesn't resolve, or points to a different paper, don't use the citation.
Does this work the same in Claude and ChatGPT?
Both connect to the same Valyu server, but setup differs and each assistant writes its own answer. Verify the citations the same way in either one.
Can I cite preprints?
You can if your field and publisher allow it, but label them as not peer reviewed and check whether a published version now exists.
Why check the registry if the paper is published?
It confirms the trial's identity and status, and it shows work that hasn't been published. Without it, a review can miss that a larger trial is still underway.
Which DeepResearch mode should I use?
Fast mode handled this one question in our test, though it stopped at its step limit. For a broader topic, try a deeper mode and check the pricing first.
Build a review you can defend
Set the boundaries, ask for the evidence table, then check every source yourself. Give your agents access to Valyu, explore the life sciences documentation, or review the academic search API comparison.
Related Blogs
How to Build a Biomedical Research Agent with PubMed and ClinicalTrials.gov
How to Find Recruiting Clinical Trials from ClinicalTrials.gov via API
The Best Academic Search APIs for AI Agents in 2026
How to Integrate PubMed Papers into Your AI (Complete 2025 Guide)
More from the blog
Valyu Add
Join 12,000+ professionals and knowledge workers.
Valyu Add is a free weekly research briefing for builders, investors and operators. Every issue is sourced, cited and verified with Valyu DeepResearch.
