live-ketcher-mcp

MCP server that connects AI assistants to a live Ketcher window, the open-source web-based chemical structure editor. Draw, edit, and organize structures directly in the canvas you have open.

Status: Tier 1 + Tier 2 underway. Transport (Phase 0) is done. Twenty-two tools are live and verified end to end through a real mcp.call_tool round trip against a real Ketcher v3.12.0 instance: ketcher_get_document_state, ketcher_diff_since_last_check, ketcher_find_duplicates, ketcher_insert_structure, ketcher_export_structure, ketcher_get_properties, ketcher_transform_structure, ketcher_check_warnings, ketcher_list_atoms, ketcher_add_atom, ketcher_edit_atom, ketcher_edit_bond, ketcher_remove, ketcher_get_stereochemistry, ketcher_set_bond_stereo, ketcher_set_enhanced_stereo, ketcher_move_structure, ketcher_get_layout, ketcher_arrange_grid, ketcher_make_arrow, ketcher_make_plus, ketcher_list_arrows. See ROADMAP.md for the full scope and tool-by-tool port map, and AGENTS.md for the transport internals and everything that only surfaced once a real server and real tools existed — including three corrections: aromatize()/dearomatize() are documented in Ketcher's own API reference but don't actually exist on this build, setMolecule() recenters the WHOLE document to a canonical viewport position on every write (confirmed: absolute coordinates are never preserved, only offsets between structures survive, exactly), and that's the actual mechanism behind the small sibling-position drift noted earlier.

Unlike its sibling project live-chemdraw-mcp, this needs no commercial license and no Windows: Ketcher is Apache 2.0 and runs in a browser.

Why this design

Ketcher exposes a real JavaScript API (getKet, setMolecule, layout, calculate, getSmiles/getMolfile/getCDXml/getInchi, and OCR via recognize). That means editing is a document round trip: read KET JSON, transform it in Python, write it back. There is no imperative object model to fight, so most tools are pure functions over JSON and are unit testable without a browser.

Three findings from a live prototype against Ketcher v3.12.0 shaped the architecture (details in ROADMAP.md section 6b):

1. setMolecule() pushes onto the undo stack rather than clearing it. AI edits are undoable, one tool call per undo entry.

2. Round trips preserve coordinates (to ~1e-6), selection, and atom ordering.

3. Writing back a subset of the document silently deletes everything omitted. The transport enforces read-all / write-all so tools cannot get this wrong.

Related MCP server: Brainstorm MCP Server

Layout

ketcher_connector/
  transport/   Playwright bridge to a live Ketcher page
    connection.py  launch/attach a visible browser running Ketcher
    document.py    read_document/apply_transform, the whole-document rule
    identity.py    stable claude-* ids via content hash, with rebind() for
                   self-initiated edits (see AGENTS.md item 1 under rule 6)
    internals.py   quarantined below-API access (undo/redo/selection)
    worker.py      dedicated thread owning the Playwright session
    convert.py     hidden scratch page: representation <-> KET fragment,
                    layout(), calculate() -- isolated from the live canvas
  bridge/      KET round-trip document surgery: get_document_state,
               diff_since_last_check, find_duplicates, insert/export_
               structure, get_properties, transform_structure
               (action='clean' only), check_warnings, list_atoms,
               add/edit_atom, edit_bond, remove, get/set_bond_stereo,
               set_enhanced_stereo, move_structure, get_layout,
               arrange_grid, make_arrow, make_plus, list_arrows
  domain/      pure logic, no editor coupling, pytest-covered
    ket_merge.py   splice/replace/centroid/remove_ref/arrow_node/
                   plus_node helpers for merging or removing a fragment
                   or annotation without disturbing the rest
    ket_atoms.py   pure atom/bond mutation (add/edit/remove/stereo),
                   including index-renumbering on removal
    ket_warnings.py   extracts valence warnings from calculate()'s own
                      error-string-in-place-of-a-number behavior
    dedup.py, diff.py, layout_math.py   ported unchanged from
      live-chemdraw-mcp -- fed KET-derived data instead of CDXML-derived
      data (layout_math's grid_positions/Box/find_overlaps power
      arrange_grid without any changes to that module at all)
    _cdxml_port_pending/   ported but CDXML-coupled, see AGENTS.md
tools/         thin MCP tool definitions
tests/         domain tests, run with no browser required
server.py      FastMCP stdio entry point
test_smoke_visual.py   live transport + bridge smoke test, real Ketcher

Setup

python -m venv .venv
.venv\Scripts\python -m pip install -e .
.venv\Scripts\python -m playwright install chromium

Register in %APPDATA%\Claude\claude_desktop_config.json:

"mcpServers": {
  "ketcher": {
    "command": "C:\\path\\to\\live-ketcher-mcp\\.venv\\Scripts\\python.exe",
    "args": ["C:\\path\\to\\live-ketcher-mcp\\server.py"]
  }
}

Restart the Claude desktop app. The server launches its own visible browser window running Ketcher — no separate install needed beyond the Playwright Chromium binary above.

Testing

python -m pytest                       # pure domain logic, no browser
python test_smoke_visual.py            # live transport check, visible window
python test_smoke_visual.py --headless # live transport check, no window

The pytest suite is pure logic against fakes and fixtures — no browser, no Ketcher instance, no network. test_smoke_visual.py is the live counterpart: it drives a real Ketcher instance and checks the properties documented in ROADMAP.md section 6b and AGENTS.md (undo behavior, whole-document write safety, identity stability) against the real app, not just against the transport code in isolation.

Provenance

The domain/ modules and their tests are ported from live-chemdraw-mcp, which is MIT licensed and by the same author. They were chosen because they carry no editor coupling: caption association, panel-box violation detection, layout math, SMARTS-based substructure matching, derivative enumeration, HRMS text generation, and IUPAC naming are all independent of which editor renders the result.

License

MIT. Ketcher itself is Apache 2.0 and is not bundled here.