Model Context Protocol Explained for Scientists

The Problem: Scientific Tools Are Hard to Use

If you are a scientist working in structural biology, computational chemistry, or drug discovery, you already know the pain. Every tool has its own installation process, its own input format, its own command-line interface, and its own way of reporting results. Predicting a protein structure requires one tool. Docking a molecule against that structure requires another. Checking whether the molecule is drug-like requires a third. Each handoff between tools means reformatting data, reading documentation, and debugging environment issues.

Now imagine you could just tell an AI assistant: “Fold this protein, dock ibuprofen against the active site, and check if ibuprofen passes Lipinski's rules.” The assistant figures out which tools to call, in what order, with the right input formats – and gives you a plain-language summary of the results. That is what the Model Context Protocol makes possible.

What Is MCP? The USB-C Analogy

The Model Context Protocol – MCP for short – is an open standard created by Anthropic that defines how AI assistants connect to external tools. The simplest way to understand it is to think about USB-C.

Before USB-C, every device had its own charging cable: micro-USB for phones, Lightning for iPhones, barrel connectors for laptops, proprietary ports for cameras. You needed a drawer full of cables and adapters. USB-C solved this by creating one universal connector that works with everything.

MCP does the same thing for AI tools. Before MCP, connecting an AI assistant to an external service required custom integration code for each tool – defining function schemas, handling authentication, parsing responses, managing errors. Every tool needed its own “cable.” MCP standardizes all of this into a single protocol. A tool server describes what it can do. An AI client reads that description. The two communicate through a well-defined message format. One standard connector for all tools.

Why MCP Matters for Science

Scientific computing workflows have a unique property that makes them especially well-suited to AI agent orchestration: they are deeply sequential, and each step requires domain expertise to interpret before proceeding to the next.

Consider a typical drug discovery workflow:

• Step 1: Predict the 3D structure of a target protein from its amino acid sequence
• Step 2: Identify potential binding pockets on the predicted structure
• Step 3: Dock candidate small molecules into those pockets
• Step 4: Calculate molecular properties and check drug-likeness rules
• Step 5: Run ADMET predictions to flag toxicity or bioavailability issues

Each step requires a different computational tool, a different input format, and domain knowledge to judge whether the results are good enough to continue. A pLDDT confidence score below 60 from protein folding might mean the structure is unreliable – proceeding to docking would be a waste of compute. A LogP above 5 might mean the compound will never be orally bioavailable, no matter how strong the binding affinity looks.

An AI assistant with access to these tools through MCP can make these judgment calls. It understands the science well enough to interpret intermediate results, decide whether to proceed, flag potential issues, and explain its reasoning. This is not replacing the scientist. It is giving them a knowledgeable computational assistant that works on demand.

How MCP Works: The Three-Step Dance

Under the hood, MCP follows a simple three-step pattern. You do not need to understand the technical details to use it, but knowing the flow helps demystify what happens when you talk to an AI assistant that has MCP tools connected.

Step 1: The Server Exposes Tools

An MCP server is a lightweight process that knows how to describe its tools and execute them. SciRouter runs an MCP server at mcp.scirouter.ai/sse that exposes over 30 scientific computing tools. Each tool has a name, a description, and a schema that defines what inputs it expects and what outputs it returns.

For example, the ESMFold tool is described with a name (“esmfold_predict”), a plain-language description (“Predict 3D protein structure from amino acid sequence”), and a parameter schema (one required field: a string called “sequence” containing single-letter amino acid codes).

Step 2: The AI Discovers Available Tools

When you open a conversation in Claude Desktop (or any MCP-compatible client), the client connects to all configured MCP servers and asks each one: “What tools do you have?” The server responds with the full catalog. The AI now knows what tools exist, what they do, and how to call them.

This discovery step is automatic. You do not need to tell the AI which tools are available or how to use them. It reads the descriptions and schemas and figures out the rest.

Step 3: The AI Calls Tools During Conversation

When you make a request that requires a tool – “fold this protein sequence” – the AI constructs the correct function call, sends it to the MCP server, waits for the result, and then interprets the output in context.

You: "Fold this sequence: MVLSPADKTNVKAAWGKVGAHAGEYGAEAL..."

Claude thinks: "This is a protein sequence. I have the ESMFold tool.
I'll call it with this sequence."

→ MCP request: esmfold_predict(sequence="MVLSPADKTNVKAAWGKVGAHAGEYGAEAL...")
← MCP response: { pdb: "...", mean_plddt: 91.2, ... }

Claude: "The structure has been predicted with high confidence
(pLDDT 91.2). It folds into a classic globin fold with 8 alpha
helices. This is consistent with human hemoglobin beta subunit..."

The key insight is that the AI does not just return raw data. It interprets the pLDDT score, recognizes the fold topology, and provides biological context – all within the same conversational turn.

Real Example: “Fold This Protein and Find Binding Pockets”

Let's walk through a concrete example of what an MCP-powered science workflow looks like from the researcher's perspective. No code. No command line. Just conversation.

You: I have a kinase I'm studying. Here's the sequence:
MENFQKVEKIGEGTYGVVYKARNKLTGEVVALKKIRLESEEEGVPSTA
Can you predict its structure and tell me about potential drug targets?

Claude: I'll fold this kinase sequence and analyze the result.

[Calls ESMFold tool — takes ~10 seconds]

The structure has been predicted with a mean pLDDT of 82.4, indicating
good overall confidence. Here's what I found:

- The fold contains the characteristic bilobal kinase domain
- The N-terminal lobe has a 5-stranded beta sheet and one alpha helix
- The C-terminal lobe is predominantly alpha-helical
- The ATP-binding cleft is located between the two lobes

For drug targeting, the ATP-binding pocket is the most common site
for kinase inhibitors. The predicted structure shows a well-defined
pocket suitable for small molecule docking.

Would you like me to dock any specific compounds into this pocket?
Or I can check molecular properties of candidate inhibitors.

The researcher typed one message. The AI called a protein folding tool, waited for results, interpreted the structure in the context of kinase biology, identified a druggable pocket, and offered next steps. The entire exchange took about 15 seconds.

SciRouter's MCP Server: 30+ Scientific Tools

SciRouter's MCP server gives AI assistants access to a growing catalog of scientific computing tools. Here are the main categories:

• Protein Structure: ESMFold for single-chain folding, Boltz-2 for multi-chain complexes, protein embeddings for sequence analysis
• Molecular Docking: DiffDock for AI-powered docking, AutoDock Vina for physics-based docking
• Molecular Properties: Molecular weight, LogP, TPSA, hydrogen bond donors and acceptors, rotatable bonds, drug-likeness checks
• ADMET Prediction: Absorption, distribution, metabolism, excretion, and toxicity profiles for drug candidates
• Format Conversion: Convert between SMILES, InChI, MOL, and PDB formats
• Similarity Search: Tanimoto fingerprint similarity to compare molecules against known drugs

Every tool is accessible through the same MCP connection. There is nothing additional to install, no separate accounts to create, and no format conversion to manage. The AI handles all of it.

Setting Up MCP: A Two-Minute Guide

Connecting SciRouter's tools to Claude Desktop requires two things: an API key and a configuration file edit. The entire process takes less than two minutes.

Get Your API Key

Sign up at scirouter.ai/register to create a free account. Go to the API Keys section of your dashboard and generate a new key. It starts with sk-sci- and is shown only once – copy it immediately.

Configure Claude Desktop

Open your Claude Desktop configuration file. On macOS it is at ~/Library/Application Support/Claude/claude_desktop_config.json. On Windows, look for %APPDATA%\Claude\claude_desktop_config.json. Add this entry:

{
  "mcpServers": {
    "scirouter": {
      "url": "https://mcp.scirouter.ai/sse",
      "headers": {
        "Authorization": "Bearer sk-sci-your-api-key"
      }
    }
  }
}

Save the file and restart Claude Desktop. You should see a hammer icon in the chat input area. Click it to confirm SciRouter's tools are listed.

Verify the Connection

Open a new conversation and ask Claude what scientific tools it has access to. It should list all the SciRouter tools it discovered. If the tools do not appear, double-check your API key and ensure there are no JSON syntax errors in the configuration file.

What About Privacy and Data Security?

Giving an AI assistant access to computational tools naturally raises questions about data security. Here is how SciRouter's MCP implementation handles the main concerns:

• Your API key stays local: The key is stored in a configuration file on your machine and sent directly to SciRouter over HTTPS. The AI model provider never sees it
• Inputs are not stored: Protein sequences and molecular structures are processed and discarded after results are returned. SciRouter does not build a database of your research queries
• Tools are read-only: SciRouter tools predict structures and calculate properties. They do not modify files, send emails, or perform any destructive operations
• Rate limits apply: MCP tool calls count against the same rate limits as direct API calls, so there are no unexpected charges

Beyond Claude: The Future of MCP in Science

MCP is still early, but the trajectory is clear. As more AI assistants adopt the protocol and more scientific tool providers expose MCP servers, the ecosystem becomes increasingly powerful. Imagine connecting to a genomics MCP server, a proteomics MCP server, and a chemistry MCP server simultaneously – then asking an AI assistant to run a complete target identification and lead optimization pipeline across all three.

The protocol also opens the door for lab automation. MCP servers can wrap not just computational tools but any system with an API – including electronic lab notebooks, sample management systems, and even robotic liquid handlers. The same conversational interface that predicts protein structures today could order reagents and schedule experiments tomorrow.

Next Steps

If you are ready to try MCP with scientific tools, here is where to go next:

• Complete MCP setup guide – detailed walkthrough with screenshots and troubleshooting
• Building an AI science agent with MCP – advanced workflows and programmatic agent construction
• Try ESMFold and DiffDock directly through SciRouter's API

Ready to connect AI to your research workflow? Get your free API key and set up MCP in under two minutes.