How to Use Consensus MCP with AI Agents for Academic Research

Consensus is an AI-powered app for researchers that provides evidence-based answers crawled from a 220M-papers-strong database. It has previously been available via the in-browser app, but the recently released MCP connector allows use from a desktop Claude Cowork or ChatGPT installation. Consensus’ capabilities can now be integrated into custom academic workflows, which opens up a great variety of use cases for academics.

This guide covers how to install Consensus MCP in Claude Cowork and ChatGPT, and some powerful academic use cases that become available through this new feature.

Video: What can you do with the Consensus MCP?

To demonstrate how powerful and flexible the Consensus MCP is, we used the Codex AI agent, gave it access to Consensus through the MCP, and asked it to create a lit review matrix. It ran 45 searches, identified 36 studies, added them to a file that can be imported into Zotero with one click, worked through each study, and identified a shortage of studies for South America, suggesting a research gap. All in about 6 minutes of work. Take a look:

While Consensus has limited access to the full-text versions of the articles, the questions we asked (main results, study area & study organism) should be largely covered in the abstract, which is always available. In 6 minutes, the agent did more work than I could have done in a few days. My goal now is to read through the details of these studies, record my findings, and investigate whether the missing South American studies really indicate a research gap!

If you think that this can be helpful, read on to understand what the MCP is, how it works and what makes it so powerful, as well as the exact prompts used for this example.

How to Install Consensus MCP

First, make sure you have an account and are signed in Consensus’ website. It does not matter if you have a paid subscription or are on the free plan; the MCP works either way (we discuss plan-related limits below).

Then, open your desktop Claude Cowork or ChatGPT app. This is important for advanced workflows. You can also use Consensus’ MCP from the in-browser Claude/ChatGPT, but these don’t have access to your local file system and cannot work with other apps (like Obsidian) on your computer.

The easiest way to do it is to just ask the AI. On either of these apps, open a new chat and ask the AI to connect to the Consensus MCP. The agent will open an authorisation page in the browser asking you to accept the MCP connection. Alternatively, you can search for Consensus on ChatGPT’s Plugins page, or on Claude’s Connectors page. All of these will still ask you to accept the connection in the same way.

Once you accept, the Consensus MCP is installed. In practice, this means you can now query Consensus directly from Claude Cowork or ChatGPT (without going to Consensus’ website).

Now that you’ve installed it, the creative part is figuring out how Consensus MCP fits into your academic workflow and understanding why it is more powerful than just using the Consensus website. Here are a few examples built around practical academic scenarios.

Why Consensus MCP can do more than the website

On the surface, the MCP is just a way of using the Consensus Website from the chat window. So what is the advantage over just using the website? It is the combination with a reasoning agent that can now use Consensus as a tool to achieve a goal you set, using the context you provide. Take a look at this diagram of the full agentic workflow, involving agentic reasoning, the Consensus MCP and external context:

To understand the benefits of each element in this diagram, let’s look at 4 consecutively expanding use cases. Below are prompts for each. By running them, you will build an understanding of how to use the Consensus MCP connector (and any other MCP server).

1. Just the MCP: First, you simply ask AI to use the Consensus website. Nothing groundbreaking, just a convenience.
2. MCP + Agent: Here, the agent uses the MCP (repeatedly) as a tool to achieve a goal you set.
3. MCP + Agent + Reasoning: Now, additionally, the agent also interprets or filters the results it receives from the MCP connector based on the output and your briefing.
4. MCP + Agent + Reasoning + Context: Finally, we add external data the agent uses via the MCP connector, enabling much better reasoning and comparison.
5. From here on, you can imagine how using MCP connectors for multiple servers makes the agent more and more powerful!

Try out the template prompts in your AI chat after setting up the MCP and experiment, then edit them to fit your needs. Turn them into skills if you want to reuse them frequently.

To get the most out of these workflows, it is also useful to understand how Consensus’ search works. There are a large number of filtering options available: publication year, study type, sample sizes, journal prestige, domain, open access, and more. The full list of filters is available in the API reference page.

1. Just the MCP: Quick fact-checking

Probably the most intuitive use of a huge scientific database at your fingertips is quickly fact-checking claims you encounter while reading or writing. Give the agent a passage and ask it to use Consensus to find evidence that supports, challenges, or qualifies it.

Use Consensus to fact-check the following claim:
[PASTE CLAIM]
Find evidence that supports, challenges, or qualifies the claim. Prioritize recent, high-quality evidence and systematic reviews or meta-analyses when available.

Duration: ~15 seconds
This use case is no different from using the Consensus website. So let’s see how having an Agent + MCP is actually different.

2. MCP + Agent: Find seminal review papers

If you have a research question but no reading list yet, ask the agent to find a small set of highly relevant review papers. Rather than relying on a single query, have the agent generate variations of the research question to improve coverage.

Use Consensus to find seminal review papers for this research question:
[RESEARCH QUESTION]
Run up to 3 different searches using alternative terminology or formulations of the question.
Select 5–10 papers that provide a strong starting point for exploring the literature. Prefer highly relevant, influential, or recent papers, while retaining important older studies when appropriate.

Duration: ~1 min
Notice that in this use case, the agent uses the MCP as a tool, which means it will run longer. It generates a research question and then uses the MCP to query. This is where the flexibility of the Agent + MCP combination becomes apparent.

3. MCP + Agent + Reasoning: Build a reading list

Consensus MCP is also useful when you already know what you want to investigate and simply need a reading list. You can harness the robust filtering system to retrieve the exact type of evidence you want, such as publication dates, sample sizes, or other screening criteria.

Use Consensus to build a short reading list on:
[TOPIC OR RESEARCH QUESTION]
Find 5–10 particularly relevant papers and use multiple queries to tackle the question from multiple angles using Consensus.
Where possible, prioritize:
- [STUDY TYPE]
- papers published after [YEAR]
- studies with at least [SAMPLE SIZE] participants

Duration: ~2 min
Similar to the previous use case, the agent uses the MCP connection as a tool, but it also evaluates, sorts, and processes the output into the format you need, so it now takes twice as long. So the MCP connection now becomes part of a larger, more complex workflow.

4. MCP + Agent + Reasoning + Context: Challenge your own research

Give the agent a draft of your (or someone else’s) work and ask it specifically to look for evidence that contradicts the core passages, exposes limitations, or suggests plausible alternative explanations. (Keep in mind that using this for peer-reviewing work is not permitted. Do not upload unpublished work without explicit permission from the authors.)

Use Consensus to challenge the following draft. 
[Attach a file or paste the work]
Break this piece down into 5-6 core claims. Actively search for evidence that could:
- contradict
- weaken
- reveal important exceptions
- suggest alternative explanations
for each claim.
Do not try to confirm my position. Prioritize the strongest counter-evidence you can find. Summarize the main challenges and cite the relevant papers using Consensus MCP.

Duration: ~3 min
In this workflow, the MCP is a tool used by the agent for review. The agent is encouraged to reason and judge (i.e. break down work into core claims) and use some other work as context. It combines all the elements we need to build more complex use cases, which again increases the time it runs.

These small workflows usually require only a handful of Consensus searches, making them useful on the free plan or whenever you want to conserve MCP usage. The duration it takes is roughly a measure of complexity. In the most complex use case, the agent took 3 minutes to complete. Agents can run for days, not just minutes. The overwhelming majority of people probably never get close to the potential available to agents. And that complexity is, at least for now, increasing exponentially.

Now let’s try to think bigger and have the agent run for 10-15 minutes at a time.

Creating a Literature Review Matrix

To demonstrate an incredibly powerful use case involving dozens of Consensus searches that require a long-running agent, we can create an AI agent that builds a literature review matrix using Consensus MCP. The results are written to a CSV file and can be added to Zotero if you have the MCP connection enabled. A literature review matrix splits the research questions along two dimensions. Typical delineations include study area, subject, or organism:

Whilst many AI-powered tools can automate lit review matrix creation, using Consensus’ MCP with a general-purpose agent lets you fully customise the output and inspect every step, improving understanding and reproducibility.

Here is the prompt to build a matrix around global climate change-related biodiversity impacts broken down by study organism and area.

Use Consensus MCP to build a literature review matrix for the research question below.
RESEARCH QUESTION
“What are the global impacts of climate change on biodiversity?”
SCREENING CRITERIA
Use the following criteria when selecting studies:
- Prefer peer-reviewed empirical studies, systematic reviews, or meta-analyses.
- Prefer recent and higher-quality evidence where possible, while retaining important older or seminal studies.
- Prefer studies with clearly reported methods, study populations/ecosystems, and outcomes.
MATRIX STRUCTURE
Rows:
- Europe
- North America
- Asia
- Africa
- South America
Columns:
- Impact on species richness
- Impact on ecosystem services
- Impact on birds
- Impact on plants
- Impact on mammals
- Impact on marine life
- Impact on freshwater life
- Range shifts / distribution changes
- Extinction risk
You may refine overly broad search wording, but do not change the substantive meaning of the rows or columns.
WORKFLOW
1. Interpret the research question and matrix structure before searching.
2. For EACH matrix cell, formulate a targeted Consensus search query. When useful, run more than one variation of a query using synonyms or more specific terminology. Save the search queries.
3. For each cell, identify 1–3 genuinely relevant studies when sufficient evidence exists.
   Prioritise:
   - direct relevance to the cell;
   - systematic reviews and meta-analyses when appropriate;
   - strong empirical studies;
   - recent evidence;
   - highly cited or methodologically strong studies where Consensus provides that information.
   Do NOT force 2–3 studies into a cell if suitable evidence cannot be found. Mark a cell as "limited evidence found" if no study met the criteria. Do not invent studies or stretch weakly related papers to fill the quota.
4. Deduplicate studies across the full matrix.
   A study may appear in multiple cells if it genuinely addresses multiple matrix dimensions, but it should only be added once to the final bibliography / Zotero import.
   Use DOI as the preferred deduplication key where available.
5. For every selected study, capture as much of the metadata provided.
6. Build a literature review matrix.
   Each cell should contain:
   - a concise 1–2 sentence synthesis of what the studies collectively report;
   - any notable disagreement or methodological difference between studies;
   - an indication when evidence is sparse.
   Avoid copy-pasting abstracts or making the cards overly large.
7. Record search provenance.
   For every matrix cell, preserve:
   - the exact Consensus query or queries used;
   - how many useful results were identified;
   - which studies were selected.
   Save this as a “Search Log” note or section so that the workflow is auditable.
OUTPUT 
1. Create a folder for this review if an appropriate location does not already exist.
2. Create:
   - one search-log/bibliography markdown file;
   - one Lit Review  CSV file
ZOTERO OUTPUT
If you have access to my local Zotero library (via mcp or otherwise):
- Add every unique included study to Zotero with all retrieved metadata.
- Do not add duplicates.
- Add them to a new collection.
If you cannot access or write to my Zotero library:
- Generate a BibTex file containing every unique included study. 
- Include all meaningful metadata retrieved for each study. Include abstracts if possible.

The agent (GPT-5.6 Terra on max effort) worked for about 6 minutes, creating a 45-cell matrix and 35 studies total from 45 targeted searches on Consensus. It saved an exhaustive audit trail as a Markdown file and added the studies to my local Zotero library. See the video above for a walk-through of the process. Here is one part of the output. Note that 6 cells were marked “limited evidence found”:

Half of the missing cells pertain to South America, suggesting a research gap in biodiversity research there. For a literature review, the studies, as well as their absence,ence can be equally important. While it is not guaranteed that there is a research gap (in fact there is a lot of biodiversity research in the Amazon), we might miss a comprehensive review on the subject. Worth investigating.

Reviewing the Literature Support For Claims Made in Your Paper

This workflow could save you a six-month round of reviews with 15 minutes of agent work. The agent identifies the claims you are making in your paper and then uses the Consensus MCP to find supporting or contrasting evidence for each of the claims. This is not unlike a peer reviewer who knows specific parts of your paper, maybe even better than you, and uses the literature to argue against what you are saying. Take a look at this short prompt:

GOAL
Provide a well-founded critique and rebuttal towards the arguments made in this paper (file attached). 
STEPS
- Identify the main claims made in this paper and their conclusions. Especially focus on the discussion section
- Query the Consensus MCP and identify papers supporting and contrasting these claims
- Create a synthesis for each claim and assign a category: supported/disputed/unsupported/novel. The novel category is for claims where no evidence can be found either way.
OUTPUT
- A document containing one section per claim (from my paper)
- Each section has subsections: a list of papers contrasting the claim, a lit supporting it, and a short synthesis based on both papers on how well this claim is supported (100w max)
- A final suggestion on which parts of the manuscript need to be addressed
- A BibTeX file containing the studies you identified (duplicates removed) with abstracts
- An overview summary of your search queries and steps 
RULES
- Pause when you reach 30 consensus searches and ask me if I want to continue, and how many claims you still have not analysed.

The output is a document summarising the claims I am making. The claims are sorted into three categories:

• Supported and Unsupported based on what the literature says
• Novel for claims where the agent could not find any evidence
• Disputed for claims that are debated in the literature

Take a look:

The supported claims usually contain additional literature I can cite to strengthen the argument. Disputed arguments need a few more sentences to convince readers, and unsupported ones you should study in detail. Be careful not to take everything as truth just because it looks professional. Remember: Consensus does not have access to all full-text PDFs. AI can make mistakes, misinterpret findings, or simply miss existing evidence. So be critical about what you find and double-check it. The studies alone, however, will give you a solid starting point for greatly improving the manuscript.

What is an AI Agent?

An AI agent is simply an AI which can sustain long-running, complicated tasks with a higher degree of autonomy. The idea of “agentic AI” has risen to mainstream popularity in the past year and a half, once desktop apps like Claude Cowork and ChatGPT (formerly “ChatGPT Codex”) broke out of the programming niche and gained more freedom to interact with your computer locally.

Whilst “old” question-and-answer AI models from 2023-2024 were mostly useful as a substitute to Google searches and Wikipedia — and struggled to perform real tasks — AI agents can use apps, connect tools and effectively engage with your entire computer to automate a lot of tedious work.

👉 Curious about AI Agents in the context of academic work? The latest AI Agents Course delves deep into how agentic AI can speed up and enhance most academic workflows.

If asked to or required by the task, AI agents can also “take a look” at the contents of your screen: see your open tabs or apps, and interact directly with the interface like a human would. In practice, however, this is a slow and wasteful way to go about it. Slow because the AI takes time to visually parse the interface and decide what to do, wasteful because in doing so, it uses a lot of “reasoning power” (which, depending on your plan, is limited at various caps) which could be better spent elsewhere.

The reality is that an AI does not need to see what is happening on the screen to carry out most tasks. The user interface is the layer through which humans interact with computers. The optimal way for AI agents to interact with an app or service is called an MCP.

What is an MCP?

The Model Context Protocol, or MCP for short, is a standard originally developed by Anthropic (the company behind Claude) for connecting an AI to a given app or service. Think of it like a sort of “instruction manual” for the AI agent. The agent reads it to understand how to use the app most effectively.

Through an MCP, the AI agent can avoid interacting with the interface and instead interact directly with the app. Saving a lot of time and reasoning at once.

Usually, each app serves its own MCP to the public and keeps it up to date so the AI agent uses the latest features. This also saves the agent additional time/effort that they would otherwise spend looking through various (possibly outdated) online sources to understand how to use an app or tool.

What is Consensus AI used for?

Consensus AI can search scientific literature, answer research questions with evidence, discover papers, and help you quickly understand what a body of research says about a topic. Unlike a general-purpose chatbot, Consensus searches its own large academic database (gathered from 250.000+ publications, with hundreds of millions of papers) and ties its answers back to actual studies.

In practice, researchers can use Consensus for:

• Exploring a YES/NO research question: get a quick overview of what the literature says about a particular scientific question.
• Finding relevant papers: discover studies that directly address a specific question or claim.
• Screening literature: use study metadata, methods, populations and outcomes to decide which papers deserve closer reading.
• Identifying possible research gaps: look for questions, populations or subtopics where the retrieved evidence is sparse or contradictory.

For a full walkthrough of the normal Consensus interface, including Deep Search, the Consensus Meter, study snapshots and Zotero-related workflows, see the complete Consensus AI review. Recently, we compared Consensus to other major AI-powered tools in the latest guide to the best AI tools for literature review in 2026.

If you want to see how Consensus fits into a broader research process, we also wrote a tutorial on how to use AI for a literature review, combining Consensus, Litmaps and Zotero into a full workflow.

The key difference between those guides and this article is that we used Consensus MCP here to let an AI agent run most of the tasks for us. The agent decided when and how to query Consensus to get the results it needed to complete the task we set for it.

Can I use Consensus MCP for free?

Yes, but with important limitations. Consensus MCP can be used by anyone, but the free tier comes with restrictions:

• 30 requests per month.
• Max 10 papers returned per request.

Additionally, on the free tier, the metadata attached to each paper lacks several useful fields (study type, AI-generated takeaway, salient excerpts, and DOI). This means Consensus’ MCP is usable on the free tier. You can ask the AI agent to retrieve information related to research questions, but keep in mind the information will be extracted from at most 10 papers per request.

The lowest-paid tier (Pro) gets more results back (up to 300 per request) with a limit of 500 requests per month. The plan goes for $20/mo (or $12/mo when committing to the yearly plan).

👉 Consensus MCP is used in the AI Agents for Academics course, which goes deeper into using agents effectively for your academic work, making you more productive and giving you a somewhat “unfair advantage”, at least until everyone else adapts. guides you through creating complex workflows involving multiple MCPs. These workflows are practical and are meant to tangibly improve your academic work.

Summary

The MCP protocol connects tools, services, or websites, like Consensus, to AI agents. The Consensus MCP is a strong choice for academics exploring AI-powered workflows and unlocks many new automations, from simple use cases to advanced, multi-step tasks that involve different tools and produce specific outputs. Given a plain text query, the Consensus MCP returns a set of papers related to that query along with some reasoning about what it matters. An AI agent can then use this information to accomplish the task you set for it. The key to using it efficiently is to think of the connection as a tool in a workflow rather than a convenience. To do this, master the 4 steps:

1. Provide the input, let AI call the MCP, analyse the input. You are just using the technical layer here.
2. Using an AI agent’s reasoning to decide on the input. Here you use the agent + technical layer and gain some autonomy.
3. Give the AI agent context (in form of papers or notes) so it can provide better input and evaluate the output. Autonomy still grows.
4. Finally, give the agent autonomy by defining the goal and providing the tools. Let the agent decide when and how to use the MCP connection.

The combination of MCP, Reasoning, Context, and Autonomy is what makes the MCP workflow so powerful. As an analogy, don’t think of a hammer (i.e. the MCP connection) as just a tool for driving nails (i.e. finding papers). Think of it as a tool to build a house. It may be just one of many tools in the workflow, but it is essential, and it enables a qualitative shift (i.e. from having no house to having a house), rather than a quantitative shift (i.e. driving nails faster).