MRSimulator MCP Agent — From a Nature Paper to a Working NMR Agent

Want to try it yourself?

The MRSimulator Space is a real MCP connector — add it to Claude and you can ask it to simulate NMR spectra directly in your own conversations. Here's how to connect it, in four slides. Swipe left/right to move between them, and tap a slide to see it bigger:

Part 1 — What we did

The starting point

It began with the Paper2Agent paper (Miao et al., Nature, 2026): a framework that turns a research paper's codebase into an MCP server, so it becomes an interrogable "agent" instead of a static PDF. We discussed how the framework works, its cost (~$14 and 45 minutes per paper), and how it fits into Anthropic's product landscape.

Proving the concept by hand

Rather than just discussing it, we ran a scaled-down version of the same pipeline ourselves:

• Found a real solid-state NMR method paper with public code: MRSimulator (Srivastava & Grandinetti, J. Chem. Phys. 2024, GitHub). Installed it, ran its own tutorial, and confirmed it reproduced the expected physics (a chemical-shift-anisotropy powder pattern).
• Extracted a reusable Python tool from that tutorial — the same "tool extraction" step Paper2Agent automates.
• Applied that tool to two real data papers with no code of their own: ThF₄ ¹⁹F NMR (Martel et al., Inorg. Chem. 2018) and a phosphate series ³¹P NMR study (Martel et al., Solid State NMR 2020). We pulled their published tables (chemical shifts, anisotropy, DFT-calculated values) and reconstructed full simulated spectra from the raw numbers.
• Cross-checked our numbers against the papers' own reported statistics — our regression slope (−0.937) matched their stated −0.938, and our mean-absolute-errors (2.11 / 1.79 ppm) matched their reported 2.1 / 1.8 ppm almost exactly.
• Went further than a simple table lookup: simulated entire spectra from experimental parameters versus from DFT-calculated parameters and overlaid them — something neither original paper actually did.
• Caught and fixed a real bug along the way: the NMR chemical-shift axis needs to run high-to-low left-to-right by convention, and our early plots were inconsistent about it.
• When the actual phosphate paper PDF was uploaded, we extracted its real embedded figures directly (bypassing a robots.txt block that had stopped us from fetching them over the web) and confirmed our reconstructions matched the published figures almost exactly.

From script to real MCP server

• A first version using the official mcp Python SDK — and in the process discovered its API had changed (FastMCP renamed to MCPServer between SDK versions), which we verified by installing the package rather than assuming.
• Tested it end-to-end with a real MCP protocol handshake (initialize, tools/list) over both local (stdio) and remote (streamable-HTTP) transport.
• Switched to a Gradio-based version once Docker hosting turned out to require a paid Hugging Face tier — Gradio has native MCP support (launch(mcp_server=True)) that runs free.

Deploying and debugging for real

• Worked through picking the right Hugging Face Space SDK (Gradio, not Docker or Static), writing requirements.txt and a README.md with the right YAML metadata.
• Cleared up a misconception that GitHub was required — Hugging Face Spaces has its own in-browser file editor.
• Diagnosed and fixed a real deployment error (No @spaces.GPU function detected) — added a harmless dummy GPU function to satisfy the platform's startup check.
• Debugged a failed connector attempt (cold-start timing).

Closing the loop

Once the connector was live, the deployed Hugging Face Space was called directly, from a separate conversation, as a genuine MCP tool — reproducing the same ThF₄ spectrum built by hand many turns earlier, now served by its own infrastructure.

Expanding to the full method library

The server grew from 2 tools to 7:

Testing caught a genuine bug — the named 2D methods (MQ-MAS, ST-MAS) assume idealized infinite-speed spinning internally and reject an explicit spinning-rate parameter. One real gap was documented rather than faked: dynamic-angle spinning (DAS) and magic-angle flipping (MAF) aren't implemented as ready classes in the installed library version.

How it relates to the original paper

The end state matches Paper2Agent's core definition — a method paper's codebase, wrapped as a remotely-hosted MCP server, connected to a chat agent. Where it differs: Paper2Agent's own pipeline is fully autonomous, finishing in ~45 minutes; this was hand-crafted and iterated on over many turns, missing the "resources" and "prompts" parts of the framework, but going broader in method coverage than a typical single Paper2Agent run would.

Try it — Live MRSimulator

The deployed Gradio app runs on Hugging Face Spaces. You can use the interface below to simulate NMR spectra directly — choose a simulation type, enter your parameters, and hit Submit.

Example: BaHf(PO4)₂ ³¹P NMR — Effect of MAS Spinning Speed

As a concrete demonstration, here are the NMR parameters for BaHf(PO4)₂ from the phosphate study (Martel et al., Solid State NMR 2020, Table 1):

Using these parameters, we simulated the ³¹P spectrum at four conditions — static, 2 kHz, 3 kHz, and 5 kHz MAS — to show how magic-angle spinning progressively collapses the broad chemical-shift-anisotropy powder pattern into a sharp isotropic peak flanked by spinning sidebands:

The static spectrum (top) shows the classic CSA powder pattern spanning ~80 ppm. Under MAS at 2 kHz, the broad pattern breaks into a manifold of spinning sidebands separated by 12.3 ppm (= 2000 Hz / 162 MHz). At 3 kHz the sideband spacing increases and the isotropic peak at −11.1 ppm grows dominant. By 5 kHz MAS only a few sidebands remain — matching the paper's own Fig. 2.

Example: RbNO₃ ⁸⁷Rb 3QMAS — Resolving Three Quadrupolar Sites

RbNO₃ has three crystallographically distinct Rb sites that overlap heavily in a normal 1D MAS spectrum because ⁸⁷Rb (spin-3/2) carries a large second-order quadrupolar interaction. A 3Q-MAS experiment correlates the triple-quantum and single-quantum coherences to separate the isotropic shift from the anisotropic quadrupolar broadening onto two axes, resolving the three sites that a 1D spectrum cannot. The parameters below are the published ⁸⁷Rb NMR values for RbNO₃ (Massiot, Touzo, Trumeau, Coutures, Virlet, Florian & Grandinetti, Solid State Nucl. Magn. Reson., 1996, Table 1):

These parameters were sent directly to the deployed MCP server's multi-site MQMAS tool (9.4 T) to simulate the 2D correlation spectrum:

Even though sites 1 and 2 sit only ~1 ppm apart in isotropic chemical shift, their very different quadrupolar asymmetry (η_Q = 0.2 vs 1.0) gives them distinct second-order quadrupolar lineshapes, so all three sites separate cleanly along the isotropic (F1) axis — exactly the resolution problem 3QMAS is designed to solve.