MolTalk — chemistry tools for AI assistants

Draw, verify and explore molecular structures with ChatGPT, Claude and other MCP clients.

MolTalk is an MCP server that gives AI assistants real chemistry tools. They can resolve compound names, validate structures, draw them inline, rotate them in 3D, analyse stereochemistry and export files for ChemDraw and other software.

MolTalk does not trust structures invented by the model. Every structure is resolved and checked with RDKit, OPSIN and a curated library of 22,000 compounds, and MolTalk fails closed: when a name is ambiguous, or stereochemistry cannot be verified, it says so instead of drawing a guess.

See it in action

Every drawing can be grabbed and rotated in 3D, right in the chat. These animations were recorded from MolTalk's own viewer (how).

Related MCP server: LocalPilot

What MolTalk can do

• Resolve names. Trivial names ("caffeine", "Wilkinson's catalyst") come from a bundled library of 22,038 compounds with about 200,000 names; systematic IUPAC names are read by OPSIN; PubChem is an optional fallback.

• Draw clean structures. 2D depictions with wedges and stereo labels, clean even for crowded molecules such as BINAP, Xantphos and porphyrins, shown inline in the chat.

• Rotate in 3D. Grab the drawing and it lifts into a 3D model that keeps the drawing's style, with lone pairs, charges and stereo labels attached to the geometry.

• Analyse stereochemistry. R/S and E/Z, atropisomers, allenes, spiro compounds and helicenes, with every descriptor checked against IUPAC's definitions, and specified, unspecified and arbitrary configurations kept apart.

• Enumerate stereoisomers, including meso forms and enantiomer pairs.

• Show electrons. Lewis-structure lone pairs, radicals and formal charges, in 2D and 3D.

• Handle metal complexes. Common catalysts and coordination compounds are assembled from database records and built in their proper geometry: square planar, tetrahedral, octahedral, sandwich.

• Export. ChemDraw (CDXML), MOL, SDF, PDB, XYZ and SMILES files, and SVG/PNG images of exactly what is on screen.

• Search substructures with SMARTS, chirality-aware.

Why MolTalk?

Language models can write SMILES, but they misremember structures, drop stereocentres and invent plausible-looking molecules. MolTalk changes what the assistant is allowed to claim:

• Structures come from data, not memory. Names are resolved by the library, OPSIN or PubChem, never generated.

• Validation is explicit. Every result reports where the structure came from (library, OPSIN, PubChem; with the PubChem CID) and its stereochemistry status: fully specified, partially specified or unspecified.

• It fails closed. Unknown names, ambiguous names ("lye": NaOH or KOH), names that do not say which stereoisomer ("glucose"), and stereodescriptors that the structure found cannot support ("(R)-BINAP" when the record has no axial stereo) are all refused, with the reason.

• Labels are checked. If the assistant draws a SMILES under a compound name, MolTalk checks the name against its library and flags a mismatch.

• Descriptors are verified, not just computed. Axial, allene and helical descriptors are measured independently on 3D coordinates and must agree with RDKit before they are shown.

Use MolTalk

The public server is

https://moltalk-411294000488.us-central1.run.app/mcp

It is free to use and needs no account or API key. Request logs record which tool was called, never the molecule (privacy policy, terms, support).

ChatGPT. With developer mode on, create an app (connector) for the URL above with No authentication, or upload the plugin package in plugin/. Enable it in a chat and ask, for example, "Using MolTalk, draw aspirin."

Claude. Settings → Connectors → Add custom connector, with the URL above. Then ask Claude to draw a molecule.

Other MCP clients. Connect to the URL with the streamable-HTTP transport, or run MolTalk locally over stdio (see Run it yourself). Clients without UI support receive the full data and can request the SVG.

Examples

Draw caffeine and identify its functional groups.
Draw (2R)-2-chloro-4-methylhexane and verify the stereochemistry.
Show the stereoisomers of BINAP.
Draw nitromethane with lone pairs.
Draw ferrocene.
Draw Wilkinson's catalyst and explain its geometry.
Draw cubane and replace four hydrogens with F, Cl, Br and I.
Export this molecule as a ChemDraw file.

More, with what to expect: examples/.

Chemistry support

Details: docs/CHEMISTRY.md.

Export

Ask for a file ("give me a ChemDraw file of that") and the viewer offers a download:

• CDXML opens directly in ChemDraw.

• MOL and SDF come in 2D (the drawing's layout) or 3D.

• PDB and XYZ come in 3D.

• SMILES is plain SMILES, or CXSMILES when coordinates are needed to state the stereochemistry.

When a format cannot carry some stereochemistry, the export says so instead of silently dropping it. The viewer's SVG and PNG buttons save exactly what is on screen, cropped, with a transparent background.

Known limitations

• 3D models are single calculated conformers for visualisation, not energy minima or measured structures.

• Trivial names outside the library need PubChem, which may be slow or busy from cloud servers.

• Hindered axes are detected by an ortho-substitution rule of thumb, not a rotation-barrier calculation.

• Multi-metal complexes, bridging ligands and η⁶-arene complexes are shown as stored. Geometry exceptions decided by ligand-field strength (NiCl₂(PPh₃)₂ is really tetrahedral) are not predicted.

• InChIKeys do not distinguish atropisomers.

• Up to 256 atoms per molecule.

The full list: docs/LIMITATIONS.md.

Run it yourself

git clone https://github.com/DomFico/moltalk.git && cd moltalk
python3 -m venv .venv && .venv/bin/pip install -e '.[dev]'
scripts/fetch-tools.sh                                  # OPSIN (needs Java 11+)
.venv/bin/moltalk                                        # stdio MCP server
.venv/bin/moltalk --transport streamable-http            # or http://127.0.0.1:8000/mcp

Or with Docker: docker build -t moltalk . && docker run --rm -p 127.0.0.1:8000:8000 -e PORT=8000 moltalk.

Deploying your own public instance on Cloud Run: docs/DEPLOYMENT.md. Development notes: docs/DEVELOPMENT.md.

Architecture

ChatGPT / Claude / MCP client
      │  MCP over HTTPS
      ▼
Cloud Run: MolTalk server (FastMCP)
      │  worker processes with time and memory limits
      ├─ names:    bundled library → OPSIN → PubChem
      ├─ RDKit:    validation, analysis, stereochemistry, 2D layout, 3D models
      └─ export:   CDXML, MOL, SDF, PDB, XYZ, SMILES
      ▼
Viewer (MCP Apps UI in the chat): drawing, 3D rotation, lone pairs, image export

More: docs/ARCHITECTURE.md · docs/MCP_AND_UI.md.

Testing

The test suite, 350+ tests, covers name resolution and validation, stereochemistry (against IUPAC reference examples), 2D layout quality, 3D geometry, metal complexes, exports, MCP behaviour and the viewer in a real browser.

.venv/bin/pip install -e '.[dev,ui-test]' && .venv/bin/pytest -q

See docs/TESTING.md.

Data sources

• RDKit for all cheminformatics.

• OPSIN for reading systematic names.

• PubChem and Wikidata for the bundled library, which is built from compounds that have a Wikipedia article, checked with RDKit and OPSIN.

See docs/DATA_SOURCES.md.

License

MolTalk is open source under the MIT License. It builds on RDKit (BSD 3-Clause) and the MCP Python SDK (MIT); OPSIN (MIT) is downloaded at setup, not included. Chemical data comes from PubChem (see NCBI's policies) and Wikidata (CC0).