Phoenix Chemistry

Phoenix — Chemistry Reasoning Agent (FutureHouse Platform)

Phoenix is the chemistry-specialized FutureHouse agent (job name MOLECULES). It is the successor to ChemCrow: an LLM agent equipped with cheminformatics tools (RDKit, retrosynthesis planners, property predictors, reaction databases) that reasons about molecules with grounded operations rather than free-form hallucination.

Use this skill when the user asks chemistry- or drug-design-flavored questions:

• Plan a synthesis route to a target molecule
• Propose new molecules satisfying constraints (Lipinski, target activity, etc.)
• Predict properties (logP, solubility, toxicity, etc.) of a SMILES
• Convert / validate / explain SMILES, InChI, IUPAC names
• Suggest reagents, solvents, conditions for a transformation

For pure literature questions about chemistry, use Crow instead. For chemistry reasoning model (open-weights), use the sibling ether0-chemistry-rewards skill.

Prerequisites

• pip install edison-client
• EDISON_API_KEY from https://platform.edisonscientific.com/profile

Minimal usage

import os
from edison_client import EdisonClient, JobNames

client = EdisonClient(api_key=os.environ["EDISON_API_KEY"])

resp = client.run_tasks_until_done({
    "name": JobNames.MOLECULES,
    "query": (
        "Propose a 4-step synthesis of imatinib from commercially available "
        "starting materials. Include reagents, conditions, and a brief "
        "rationale for each step."
    ),
})
print(resp.answer)

Recipes

Retrosynthesis plan

resp = client.run_tasks_until_done({
    "name": JobNames.MOLECULES,
    "query": "Plan a retrosynthesis for SMILES Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1Nc1nccc(-c2cccnc2)n1",
})

Lead optimization brainstorm

resp = client.run_tasks_until_done({
    "name": JobNames.MOLECULES,
    "query": (
        "Starting from aspirin (CC(=O)Oc1ccccc1C(=O)O), propose 5 analogs that "
        "preserve COX inhibition but reduce gastric side effects. Briefly "
        "justify each modification."
    ),
})

Property check before ordering compounds

resp = client.run_tasks_until_done({
    "name": JobNames.MOLECULES,
    "query": (
        "For each of these candidate kinase inhibitors, compute logP, MW, "
        "TPSA, HBD/HBA, rotatable bonds, and rule-of-5 compliance. Flag any "
        "that fail Lipinski:\n"
        "1. CC(C)c1nc(...)\n"
        "2. ..."
    ),
})

Reaction feasibility / condition lookup

resp = client.run_tasks_until_done({
    "name": JobNames.MOLECULES,
    "query": (
        "I want to do a Suzuki coupling between 4-bromopyridine and "
        "phenylboronic acid. What ligand, base, solvent, and temperature "
        "would you recommend, and what side products should I watch for?"
    ),
})

Demo-friendly examples (good first calls)

• "What is the IUPAC name of CN1C=NC2=C1C(=O)N(C)C(=O)N2C?" → caffeine
• "Suggest 3 isosteres of a tert-butyl group that would maintain steric bulk but reduce metabolic liability."
• "Predict the logP of ibuprofen from its SMILES and explain the contributing groups."

Cost / latency

• ~30 s – 4 min depending on whether retrosynthesis tools are invoked
• Credits: typically between Crow and Falcon

Failure modes / caveats

• Phoenix's predictions are model-grounded but not infallible — for safety-critical chemistry (toxicology, regulatory) treat outputs as a hypothesis to verify, not a final answer.
• Very long SMILES with stereochemistry should be sanity-checked with RDKit locally.
• Patent-protected synthesis routes may be missing from the underlying tools' databases.