Exposes the on-prem Flavormancer models as Model-Context-Protocol tools, so any MCP client (Claude Desktop, an agent, another app) can call the trained taste/aroma models to read a molecule, analyze a whole formulation, screen a mixture, predict reaction products, find substitutes, explore stereoisomers, search by note or flavor, browse the library, and query the flavor-space map.
Flavormancer has 16 well-named, verb-prefixed tools with consistent descriptions (100-200 chars each). All tools have input schemas with typed parameters and descriptions. However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. Error handling is minimal (generic HTTP status codes, no recovery guidance). No tool annotations (readOnlyHint, destructiveHint). Parameter descriptions lack format constraints (e.g., SMILES validation rules, k bounds). The _strip() helper removes SVG/PNG blobs, but this is implementation detail, not schema documentation. Overall: solid naming and parameter coverage, but missing output documentation and error guidance that would push this to 75+.
Analyze a formulation (list of ingredients with optional target notes and processes).
Browse molecules in a flavor category.
Design a recipe based on target flavors and aroma notes.
Look up molecules by flavor name.
Search for molecules by aroma notes.
Find structural neighbors (Tanimoto/Morgan nearest neighbors) of a molecule.
Find substitute molecules based on taste+aroma profile similarity (profile-based swaps).
Output schemas not documented. LLMs cannot predict response structure (fields, types, nesting) for downstream tool chaining. E.g., read_flavor returns taste_probabilities dict, but schema is invisible to the agent.
Error handling is generic. Tools return {"error": "HTTP 400"} with no recovery guidance. LLMs cannot determine if error is retryable, user-fixable, or fatal. E.g., 'could not read molecule' gives no hint to try search_by_flavor or validate SMILES.
Parameter constraints missing. 'k' (number of neighbors/substitutes) has no min/max bounds documented. SMILES parameter has no format validation hint. Descriptions lack actionable constraints (e.g., 'k: 1-100', 'SMILES: valid RDKit SMILES string').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Query the flavor-space map (2D/3D UMAP embedding by structure).
Interpret a natural language request about flavors.
List all flavor categories available in the library.
List stereoisomers of a molecule. The trained models are achiral (2D fingerprints), so any odor/taste difference between stereoisomers is documented, not predicted.
Find single molecules equivalent to a mixture based on taste+aroma profile matching.
Predict reactions and hazards from a mixture of ingredients and processes. Template-based and indicative — the functional groups to form these are present; this is not a claim the reaction proceeds, nor a yield/stability assay.
Read a single molecule's flavor profile — taste probabilities, confident aromas, mouthfeel scores, GRAS status, structural alerts, and tox flags.
The complete read — everything /api/predict returns (physchem, stability, chemesthesis, chirality, analytical, labeling, references, safety) plus names + full aroma scores.
Screen a mixture of ingredients for safety and flavor interactions.
No tool annotations. All 16 tools are read-only (risk: READ_ONLY), but readOnlyHint is not set in schema. Agents cannot infer safety/idempotency without explicit annotations.
Pagination not documented. Tools like browse_category and flavor_map accept 'limit' but no 'offset'/'cursor' or total count returned. Large result sets risk context window exhaustion.