IntegrationsAI Agents

Claude for Science: How AI Can Run Real Lab Workflows

Use Claude's tool-use capabilities with SciRouter MCP to fold proteins, dock molecules, and screen compounds — all through natural language conversation.

Ryan Bethencourt
April 28, 2026
10 min read

Claude Meets the Lab Bench

Claude is one of the most capable AI assistants available, with strong reasoning abilities, nuanced understanding of scientific literature, and the ability to engage in extended multi-turn conversations about complex topics. But until recently, Claude could only talk about science — it could not do science. Ask Claude to fold a protein, and it would describe the process eloquently while being unable to actually run the computation.

That changed with tool use and MCP (Model Context Protocol). When connected to SciRouter's scientific computing tools via MCP, Claude can call real computational biology and chemistry tools during a conversation. Instead of approximating a molecular weight from memory, Claude calls RDKit. Instead of describing what a protein structure might look like, Claude calls ESMFold and shows you the actual prediction with confidence scores.

This transforms Claude from a science communicator into a science operator — an AI assistant that can actually run lab workflows, interpret real computational results, and help researchers make data-driven decisions.

How Claude's Tool Use Works

Claude's tool-use capability follows a simple loop: the user asks a question, Claude decides whether it needs external computation, calls the appropriate tool, receives results, and incorporates them into its response. This happens transparently within the conversation.

With SciRouter's MCP server, Claude automatically discovers all available scientific tools when the conversation starts. Claude knows what each tool does, what parameters it accepts, and what results it returns. When a user's request requires computation, Claude selects the right tool without being told which one to use.

MCP: The Connection Layer

MCP is an open standard that lets AI assistants connect to external tool servers. Think of it as a USB port for AI — plug in a tool server and the AI can use whatever tools the server provides. SciRouter's MCP server exposes 30+ scientific computing tools including protein folding, molecular docking, property calculation, ADMET prediction, and more.

Setting up the connection requires adding SciRouter's MCP server to your Claude Desktop configuration:

claude_desktop_config.json
{
  "mcpServers": {
    "scirouter": {
      "url": "https://mcp.scirouter.ai/sse",
      "headers": {
        "Authorization": "Bearer sk-sci-YOUR_API_KEY"
      }
    }
  }
}

After restarting Claude Desktop, all SciRouter tools become available in your conversations. No additional setup, no Python environments, no GPU configuration.

Real Conversation Examples

Here are examples of the kind of scientific workflows Claude can run when connected to SciRouter tools. These represent actual tool-call patterns, not hypothetical capabilities.

Example 1: Fold a Protein and Assess Quality

Conversation: Protein folding
User: Fold this protein and tell me about the structure:
      FVNQHLCGSHLVEALYLVCGERGFFYTPKT

Claude: I'll fold that sequence using ESMFold.
[Calls: ESMFold with sequence FVNQHLCGSHLVEALYLVCGERGFFYTPKT]

Claude: Your protein folded successfully. Key results:
- Residues: 30
- Mean pLDDT: 87.3 (high confidence)
- Structure: Primarily alpha-helical with a beta turn
- Notable: Predicted disulfide bond between Cys6 and Cys11

This is the B-chain of human insulin. The high pLDDT
indicates a well-folded, structured peptide. The disulfide
bond is consistent with known insulin biochemistry.

Example 2: Dock a Drug Candidate

Conversation: Molecular docking
User: Can you dock imatinib against ABL kinase and tell me
      about the binding?

Claude: I'll dock imatinib against ABL kinase using DiffDock.
[Calls: DiffDock with imatinib SMILES and ABL kinase PDB]

Claude: Docking complete. Here are the results:
- Top binding pose confidence: 0.89
- Predicted binding energy: -9.4 kcal/mol
- Key interactions: H-bond with Thr315 gatekeeper,
  hydrophobic contacts with the C-helix

The strong predicted binding is consistent with imatinib's
known activity as a BCR-ABL inhibitor. The interaction
with Thr315 is clinically relevant — the T315I mutation
(Thr to Ile) is the most common resistance mechanism.

Example 3: Screen a Compound for Drug-Likeness

Conversation: Compound profiling
User: Is this molecule drug-like? CC(=O)Oc1ccccc1C(=O)O

Claude: Let me analyze that compound's properties.
[Calls: Molecular Properties with SMILES]
[Calls: ADMET Prediction with SMILES]

Claude: This is aspirin (acetylsalicylic acid). Property profile:
- MW: 180.16 (passes Lipinski, < 500)
- LogP: 1.2 (passes Lipinski, < 5)
- H-bond donors: 1 (passes, < 5)
- H-bond acceptors: 4 (passes, < 10)
- TPSA: 63.6 Ų (good oral absorption)
- Lipinski violations: 0

ADMET predictions:
- Absorption: High (0.92)
- Toxicity: Low risk
- Metabolized primarily by esterases

Verdict: Excellent drug-likeness profile. All Lipinski rules
pass with comfortable margins.

Multi-Step Workflows: Where Claude Shines

Individual tool calls are useful, but the real power emerges when Claude chains multiple tools together in a single conversation. This is where Claude's reasoning abilities combine with SciRouter's computational tools to create something neither could do alone.

Consider this workflow: "I have a new protein target. Can you predict its structure, find druggable pockets, and then dock aspirin against the best pocket?" Claude will:

  • Call ESMFold to predict the protein structure
  • Call Pocket Detection on the predicted structure
  • Analyze the pockets and select the most druggable one
  • Call DiffDock to dock aspirin against that pocket
  • Interpret all results and provide a coherent summary

Each step builds on the previous one. Claude makes decisions about which pocket to target based on the detection results. It selects the right docking parameters based on the protein structure. And it synthesizes everything into a narrative that a researcher can understand and act on.

Setting Up Claude Desktop with SciRouter

Getting started takes about five minutes:

  • Create a SciRouter account at scirouter.ai/register (free tier includes 500 credits)
  • Copy your API key from the dashboard
  • Open Claude Desktop settings and navigate to the MCP configuration
  • Add the SciRouter MCP server with your API key
  • Restart Claude Desktop

After setup, start a new conversation and ask Claude to fold a protein or analyze a molecule. You will see tool-call indicators in the conversation showing which SciRouter tools Claude is using, along with the computed results.

Note
For detailed setup instructions with screenshots, see our complete MCP setup guide.

Limitations and Responsible Use

Claude connected to SciRouter tools is a powerful research assistant, but it has important limitations that researchers should understand:

Claude Is Not a Domain Expert

Claude has broad knowledge of biology and chemistry but can make reasoning errors, especially in specialized subfields. It may misinterpret an unusual pLDDT pattern or draw incorrect conclusions from docking results. Always review Claude's interpretations with domain expertise, especially for results that will influence experimental decisions.

Computational Predictions Are Approximations

ESMFold predictions are not crystal structures. Docking scores are not binding affinities. ADMET predictions have accuracy limits. These tools narrow the search space and prioritize candidates, but they do not replace experimental validation. Claude should help you decide what experiments to run, not whether to skip experiments entirely.

Verify Critical Results

For any result that will influence significant research decisions — compound selection, target validation, clinical strategy — verify the computational predictions with independent methods. Run the same analysis through a different tool, check results against published data, and consult domain experts. Claude is a starting point, not an endpoint.

What Comes Next

The combination of Claude's reasoning with SciRouter's computation is just the beginning. As MCP becomes more widely adopted, we expect to see researchers routinely using AI assistants for computational science — not replacing their expertise, but amplifying it. A researcher who can run protein folding, docking, and ADMET analysis through natural conversation can explore more hypotheses in less time, catch problems earlier, and make better-informed decisions about where to invest lab resources.

SciRouter is expanding its tool catalog continuously. New tools for molecular generation, retrosynthesis, crystal structure prediction, and genomics analysis are joining the platform. Each new tool automatically becomes available through MCP, meaning Claude's scientific capabilities grow with every SciRouter update.

Tip
Try it yourself: create a free SciRouter account, set up MCP in Claude Desktop, and ask Claude to fold the insulin B-chain (FVNQHLCGSHLVEALYLVCGERGFFYTPKT). In under a minute, you will have a predicted structure with confidence scores, interpreted by an AI that understands what the results mean.

Frequently Asked Questions

Can Claude actually fold proteins?

Claude itself does not fold proteins — it connects to ESMFold via SciRouter's MCP server to run the actual computation. Claude decides when to call the tool, sends the sequence, receives the predicted structure, and interprets the results in natural language. The protein folding happens on SciRouter's GPU infrastructure; Claude provides the reasoning and communication layer.

How do I set up Claude Desktop with SciRouter MCP?

Add SciRouter's MCP server to your Claude Desktop configuration file (claude_desktop_config.json). Specify the server URL (mcp.scirouter.ai) and your SciRouter API key. Restart Claude Desktop, and all 30+ scientific tools will be automatically available. See our MCP setup guide for step-by-step instructions.

Is Claude's scientific reasoning reliable?

Claude's reasoning about scientific concepts is generally strong but not infallible. When connected to SciRouter tools, the key reliability improvement is that computational results come from validated tools (ESMFold, RDKit, DiffDock) rather than from the LLM's memory. Claude may occasionally misinterpret results, so domain expertise remains valuable for reviewing conclusions.

What is the cost of running science workflows through Claude?

You pay for two things: your Claude subscription (for the conversation) and SciRouter credits (for the computation). A protein fold costs 5 SciRouter credits. Docking costs 10 credits. Property calculations cost 1 credit. SciRouter's free tier includes 500 credits per month. Claude's cost depends on your subscription plan with Anthropic.

Can Claude chain multiple tools in a single conversation?

Yes. Claude can call multiple SciRouter tools in sequence within one conversation. For example, you can ask Claude to fold a protein, find its binding pockets, and dock a drug candidate — all in a single conversation thread. Each tool call builds on the previous results, enabling complex multi-step workflows through natural language.

Try this yourself

500 free credits. No credit card required.