ChemGlyph

ChemGlyph

Publication-quality chemical structure and reaction rendering for AI agents. ChemGlyph is the KaTeX of chemistry: a rendering layer, a validation layer, and an MCP interface on top of RDKit.

Install

pip install chemglyph

Render a molecule

import chemglyph

result = chemglyph.render_molecule("c1ccccc1")  # benzene
open("benzene.svg", "w").write(result.data)

render_molecule takes SMILES, InChI, or molblock and returns SVG (or PNG) plus canonical_smiles, mol_formula, mol_weight, and warnings.

Styles

Three styles, same molecules (benzoic acid, caffeine, (S)-ibuprofen):

chemglyph.render_molecule(smiles)  # screen: coloured O, N, S, Cl for screens and chat
chemglyph.render_molecule(smiles, style="acs1996")  # black, ACS journal sizes
chemglyph.render_molecule(smiles, style="textbook")  # heavier lines for teaching

Every style draws at one physical scale. acs1996 follows the ACS Document 1996 settings: 14.4 pt bonds, 0.6 pt lines, 10 pt labels, and bond ends that stop 1.6 pt from each label. Two structures drawn with the same style have the same bond length and label size, so they line up in a paper. screen keeps that geometry and colours O, N, S, and Cl (classic CPK); textbook uses 0.8 pt lines and 12 pt labels.

The SVG size is given in points and the background is transparent unless transparent=False. fmt="png" takes a dpi, fmt="pdf" gives a vector PDF for LaTeX, and fmt="mol" a molfile with the drawn coordinates for a structure editor.

result = chemglyph.render_molecule("CC(=O)Oc1ccccc1C(=O)O", style="acs1996")
result.report  # {'bond_length_pt': [14.4, 14.4], 'label_clearance_pt': 1.6, ...}
png = chemglyph.render_molecule(smiles, style="acs1996", fmt="png", dpi=600)

The style names from 0.1 (acs, modern, textbook-cn) still work, map to acs1996, screen, and textbook, and give a deprecation warning; they go in 1.0.

These numbers are measured in the rendered output by the test suite (tests/test_profiles.py).

Compound numbers

chemglyph.render_molecule(smiles, style="acs1996", number="12a")

The number is set in bold at the label size, centred half a bond length under the structure, as in ACS journals. It is also stored with the embedded structure (the molfile title line), so a figure still says which compound it shows when it is read back.

Abbreviations

chemglyph.render_molecule(smiles, abbreviate=True)  # Boc, Ts, TBS, OAc, NHBoc, ...
chemglyph.render_molecule(smiles, abbreviate=["Ph", "t-Bu", "CO2Me"])

abbreviate=True draws protecting and activating groups as labels, with the spellings of the J. Org. Chem. list of standard abbreviations. Acyl, carbamate, and benzyl groups are abbreviated only on heteroatoms, so a benzyl side chain stays a CH2 and a ring; silyl and sulfonyl groups also on carbon. A group on a terminal O or NH joins its label (OTs, BocHN). A label that would run into another label or a bond is drawn in full instead.

Only the drawing changes. The embedded structure, the identity, and the name check use every atom, and fmt="mol" keeps the abbreviations as SUP groups, which structure editors show as labels and can expand.

Checked figures

Figures from these profiles check themselves and carry their structure:

result = chemglyph.render_molecule(smiles, style="acs1996", name="cholesterol")
result.report["verified"]  # bonds and labels drawn, wedges right, structure reads back
result.report["name_check"]["status"]  # match, stereo_differs, mismatch, or unchecked
chemglyph.read_structure(result.data)  # the RDKit molecule embedded in the SVG or PNG

A name that belongs to another structure produces a warning, which catches a wrong SMILES from an AI model before the figure reaches a manuscript. Names resolve offline through the built-in Chinese dictionary or OPSIN.

Reactions

spec = {
    "steps": [
        {
            "reactants": ["OC(=O)c1ccccc1O", "CC(=O)OC(C)=O"],
            "products": ["CC(=O)Oc1ccccc1C(=O)O", "CC(=O)O"],
            "conditions": {"above": "H₂SO₄ (cat.)", "below": "rt, 15 min"},
            "yield": "89%",
            "arrow": "forward",
        }
    ],
    "style": "screen",
}
svg = chemglyph.render_reaction(spec)

Reaction schemes are not yet drawn at a fixed scale; that comes with the reaction phase of the roadmap.

Conditions are pre-formatted Unicode text, so pass H₂SO₄, not H2SO4. ChemGlyph does not parse formulas out of text. The full schema (multi-step chains, equilibrium and retro arrows, line wrapping) is in docs/reaction_schema.md.

The aspirin demo writes a two-step route:

python examples/aspirin_synthesis.py  # writes examples/aspirin_synthesis.svg

Validation

validate_structure reports parse errors and applies four quick fixes: unmatched brackets and ring closures (reported, not guessed), kekulization failures of lowercase aromatic atoms, and nitrogen valence errors via a formal [N+]. Anything else passes RDKit's message through unchanged.

report = chemglyph.validate_structure("c1cccc1")
report.fixes[0].description  # 'lowercase aromatic atoms could not be kekulized...'
report.fixes[0].fixed_smiles  # 'C1CCCC1'

Naming

chemglyph.parse_name("aspirin")  # 'CC(=O)Oc1ccccc1C(=O)O'

English IUPAC and common names resolve offline through OPSIN (pip install 'chemglyph[opsin]', plus a Java runtime). Chinese names use the built-in dictionary, and the library API accepts a translator callable for names that are not in it:

chemglyph.parse_name("阿司匹林")  # 'CC(=O)Oc1ccccc1C(=O)O'
chemglyph.parse_name("六甲基苯", translator=to_english)

ChemGlyph itself never calls an online service, including for translation.

MCP server

Run the bundled console script (stdio transport):

chemglyph-mcp

Claude Desktop registration (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "chemglyph": {
      "command": "chemglyph-mcp"
    }
  }
}

Tool	Use it when	Returns
render_molecule	the user asks to draw one structure from SMILES/InChI/molblock	PNG image plus formula, MW, warnings (SVG source on request)
render_reaction	the user asks for a reaction or synthesis route	PNG image of the reaction scheme
validate_structure	a SMILES may be malformed and you need a repair	validation report JSON
parse_name	the user gives a name like "aspirin" instead of SMILES	canonical SMILES or an error

One thing to know about clients. Some MCP clients, LM Studio included, only pass the text part of a tool result to the model and never display the attached image. The render tools write their PNG to ~/Downloads/chemglyph/ and return that path in the text, so you can always open the file yourself. If a model claims it rendered a figure but nothing shows up, ask it for the saved path rather than having it redraw the structure by hand.

Benchmarks

benchmarks/ holds the fixed 20-molecule blind test and a generator that writes shuffled, numbered PNG/SVG figures plus answer_key.json:

python benchmarks/generate_blind_test.py --seed 1234

The protocol, the A/B deck generator, the runbook, and the scorer are all in benchmarks/; the pass threshold and procedure are in benchmarks/RUNBOOK.md. Anyone can run it and contribute results.

No blind test against ChemDraw has been run yet, so no comparison is claimed. Every render is instead checked as it is made (see "Checked figures"), and tests/test_corpus.py renders a 416-structure corpus on every commit.

Known limitations

• 14 of the 416 corpus structures have a layout problem that is reported with a warning: helicene and fullerene (no planar drawing); BINAP, morphine, codeine, paclitaxel, vinblastine, tetrodotoxin (crowded polycycles); and cyclodextrins, amphotericin B, forskolin, rifampicin, sitagliptin (labels in crowded places). Passing a molfile with your own coordinates avoids them. The list is in tests/test_corpus.py.
• Reaction schemes still use the 0.1 renderer, not the fixed scale.
• Full automatic Chinese name-to-structure parsing is not implemented; Chinese names resolve through a small built-in dictionary plus an optional translator hook. English names resolve through OPSIN.

Roadmap

Work proceeds in phases, and each phase ships as a minor version once its exit criteria pass. The plan, the standards ChemGlyph follows, and the quality gates are in docs/SPEC.md; release notes are in CHANGELOG.md.

• 0.1.3: fixes and release hygiene.
• 0.2 (released): molecules done right. One fixed physical scale for every drawing, style profiles measured against published standards (ACS Document 1996 settings first), a verification report with every render, automated layout checks, abbreviations, compound numbers, PDF and MOL output.
• 0.3: formulas and equations in mhchem syntax, with atom and charge balance checks.
• 0.4: reaction schemes on the shared scale, with an object API for compounds and figures.
• 0.5: mechanism (curved) arrows anchored to atoms and bonds.

Non-goals

No structure editor GUI (Ketcher/ChemDraw competition), no 3D visualization, no retrosynthesis or property prediction, no online database queries, and no automatic mechanism generation. The full list is in the project specification.

Development

python -m venv .venv
.venv/bin/pip install -e ".[dev]"
.venv/bin/ruff check . && .venv/bin/ruff format . && .venv/bin/pytest

Python 3.11+, RDKit 2024.9+, MIT license. All errors derive from chemglyph.errors.ChemGlyphError.