Materials AI is most useful when each component has a bounded job: retrieve records, prepare calculations, analyze artifacts, or develop hypotheses. A conversational interface does not make those jobs scientifically interchangeable. Start with the research question and the evidence needed to answer it, then choose the smallest combination of resources that can support that work.

This guide covers materials discovery, computational chemistry and scientific-data workflows. The combinations below are proposed evaluation designs, not tested interoperability claims.

1. Understand the layers before choosing tools

Materials Project is the upstream data service; mp-api is a client for accessing it. The third-party Materials Project MCP server exposes database operations to assistant clients. It is neither a new materials database nor a predictive model.

Agents such as ChatMOF and SciAgents coordinate reasoning and tools within particular research systems. Skills are procedural instructions, sometimes accompanied by helpers. The Pymatgen, ASE, Phonopy and pycalphad Skills described here are distinct from their upstream scientific packages. Loading a Skill does not install a calculator, obtain a thermodynamic database or provision compute resources. Host support for Skills also differs.

For the architectural distinction, see MCP, Skills, agents and APIs. The MCP in chemistry guide provides protocol context, while AI Skills for chemistry explains instruction-based workflows. Current protocol context does not establish compatibility for an older project's client configuration.

For the orchestration layer, see AI Agents for Chemistry.

2. Match each resource to a research responsibility

The following roles reflect the supplied documentation. They are not a prescription to connect every resource into one pipeline.

Resource Documented role Main inputs and outputs Important boundary
Materials Project MCP Database access through mp_api Element/property filters or identifiers → records, structures and selected property data Third-party wrapper; needs Materials Project authentication
Pymatgen Skill Structure validation, conversion planning, symmetry sensitivity and local hull analysis Structures or compatible energy entries → reports, converted artifacts and phase-diagram analyses Computed hull depends on energies and competing phases supplied
ASE Skill Route workflow preparation separately from calculator configuration Task intent → branch, rationale, missing inputs and next delegation Top-level router does not execute calculations
Phonopy Skill Organize displacement/force datasets and phonon analysis Structure, settings and forces → assembly status and phonon-band/DOS/thermal files Separate force provider and submission workflow required
pycalphad Skill Guide local finite-temperature TDB equilibrium calculations Database and composition/condition settings → numerical report and phase-equilibria CSV Equilibrium and numerical consistency are not kinetic or experimental validation
ChatMOF MOF-focused retrieval, prediction and generation Natural-language request → task-dependent information, predictions or structures Local prediction/generation need additional modules; generation needs GRIDAY
SciAgents Graph-guided hypothesis development and critique Keywords or graph exploration → structured hypotheses and research drafts Proposals and novelty assessments need independent evaluation

3. Begin with a bounded database question

The Materials Project MCP README documents search_materials searches by elements, band-gap range and stability, plus structure retrieval through get_structure_by_id. Other tools cover electronic and phonon data, elasticity, dielectric properties and additional materials records. An advertised tool does not guarantee that a particular candidate has the requested data.

Before retrieval, propose an explicit query contract: chemical system, allowed elements, requested fields, result limit and how missing values will be represented. Separate “not returned” from “calculated to be absent.” Retain identifiers, retrieval time, query criteria and available calculation origins.

The Pymatgen Skill offers an alternative bounded retrieval route: its Materials Project helper defaults to offline planning, with network execution requiring explicit approval and authentication. Choose one retrieval route initially rather than duplicating requests through two interfaces without a reconciliation plan.

Decision point: if the available records cannot support the intended comparison, narrow the question or plan new calculations. Do not fill missing properties with assistant-generated estimates presented as database values.

4. Proposed example: screen crystals, then investigate selected candidates

Consider a proposed Li–Fe–O crystal-screening exercise. Its objective is to assemble a traceable shortlist for further computation, not announce a newly validated material.

Inputs: a declared chemical-system scope, user-selected screening criteria, permission for bounded retrieval and a statement of which computed properties matter. No candidate results or performance outcomes are assumed here.

  1. Retrieve a bounded candidate set. Use the MCP server or the Pymatgen query helper to obtain identifiers, structures and available thermodynamic metadata. Save the returned records unchanged alongside an analysis copy.
  2. Validate structures. Use the Pymatgen Skill's intake workflow to inspect lattice, periodicity, coordinate convention, occupancies, parser warnings and oxidation-state treatment. Generate a symmetry-sensitivity report rather than relying on one unexplained tolerance.
  3. Review the shortlist. Separate candidates excluded by stated criteria from records deferred because of missing or ambiguous data. Preserve the reason for each decision.
  4. Prepare optional new calculations. For candidates requiring relaxation or static calculations, use the ASE router to identify the workflow branch and calculator-configuration dependency. A researcher must still select an appropriate backend and execution environment.
  5. Analyze evidence appropriate to the question. Build a local hull only with compatible energies and sufficient competing phases. For a phonon question, establish displaced-supercell forces or precomputed force constants and a traceable assembly plan before analysis.

Outputs: a candidate table, validation reports, a calculation-preparation handoff, optional analysis artifacts and a manifest connecting each artifact to its source.

Stop/go decisions: unresolved disorder can block conversion; incompatible energy conventions block hull comparison; missing displacement forces block phonon assembly. A candidate may be deferred rather than accepted or rejected when the evidence is incomplete.

5. Keep computation and analysis handoffs explicit

A proposed architecture is: bounded retrieval → structure intake → human-approved calculation plan → separately executed calculations → analysis → evidence report. Each arrow is a handoff to evaluate, not a proven connector.

The ASE Skill routes static, relaxation, molecular-dynamics and NEB requests to ase/ase-workflows, and backend configuration to ase/ase-calculators. Mixed requests start with the workflow branch. Its required response identifies the selected branch, rationale, missing inputs and next delegation. Requested execution is delegated through dpdisp-submit; detailed downstream behavior is not established by the supplied top-level file.

The Phonopy Skill separates force evaluation from displacement generation and analysis. Collect supercell, displacement amplitude, primitive-cell context and symmetry choices. Phonon bands require a defined path and sampling; DOS and thermal outputs require mesh settings, with temperatures for thermal analysis. Preserve displacement-to-force-file correspondence before assembling FORCE_SETS or force constants.

A proposed ASE-to-Phonopy handoff should therefore verify cell conventions, species order, force units and file correspondence. Shared scientific subject matter alone is not evidence that two instruction resources interoperate.

6. Distinguish four meanings of stability

Convex-hull screening: the Pymatgen Skill's local phase-diagram analysis compares supplied computed entries. Hull membership is conditional on the competing phases, compatible total energies and correction scheme. It is not a direct statement of experimental stability.

Finite-temperature equilibrium: the pycalphad Skill uses a local TDB, selected components/phases, bulk elemental mole fractions, pressure and temperatures. Its helper requires exactly N-1 independent elemental fractions and a dependent non-vacancy element. It exports molar phase fractions and compositions, retaining separate composition sets with the same phase name.

Phonon assessment: interpretation depends on force-data quality, supercell and convergence choices, units and relevant long-range corrections. The Phonopy source explicitly identifies imaginary modes arising from insufficient convergence or setup choices; treat such outputs as a diagnostic question, not an automatic experimental verdict.

Experimental stability: none of these workflows establishes it. CALPHAD equilibrium does not predict precipitation rates or retained metastable microstructures, and numerical checks do not establish database accuracy.

Use pycalphad as a separate branch when a suitable thermodynamic database supports the question. Its bundled ideal-cu-ni.tdb is a hypothetical teaching model, not an assessed Cu–Ni database or a substitute for one in the crystal-screening example.

7. Use agents where their specialization fits

ChatMOF is appropriate to evaluate when the question concerns MOFs and natural-language retrieval, prediction or structure generation. Its architecture assigns planning/tool selection to an agent, task work to toolkits and response assembly to an evaluator. That evaluator is not evidence of scientific validation. The online demonstration is search-focused; the README warns that prediction and generation generally do not work there beyond supplied examples.

SciAgents occupies a different stage: developing bio-inspired research hypotheses from scientific knowledge graphs. Its Ontologist clarifies concepts, Scientist roles develop proposals and a Critic examines weaknesses. The automated approach adds planning and novelty checking. Running it requires GraphReasoning, APIs, graph files and embeddings rather than just an LLM conversation.

Recommended proposed combinations are deliberately small: retrieval plus Pymatgen for database triage; structure intake plus ASE preparation for new calculations; a force-provider workflow plus Phonopy for phonon analysis; pycalphad plus an appropriate TDB for equilibrium questions. Use ChatMOF for MOF-specific tasks and SciAgents for hypothesis exploration, not as mandatory stages of inorganic crystal screening.

8. Handle failures and deployment uncertainty visibly

For access failures, inspect authentication and deployment before changing scientific filters. The MCP source documents Docker and local Python routes, with Python 3.12 or later and uv for local setup, and client examples for Claude Desktop and VS Code Copilot. Evaluate these configurations in the intended host; documentation is not a compatibility test.

For malformed structures or lossy conversions, preserve the original and report warnings. The Pymatgen validator's distance checks are not complete for contacts across very short lattice vectors. For incomplete force datasets, stop assembly. For failed pycalphad mass-balance or sampling-sensitivity checks, retain the failures in the report rather than relabeling the result converged. Inspect solver-imposed composition adjustments near endpoints.

Rights also need artifact-specific review. SciAgents has conflicting Apache/MIT code declarations, and the ASE Skill's MIT declaration differs from its collection's LGPL text. Clarify applicable terms before redistribution. Repository availability does not settle those conflicts or establish rights to external databases, graphs or model assets.

9. Reader checklist

  • Define the research question and evidence needed before selecting tools.
  • Separate database values, predictions, hypotheses and new calculations.
  • Bound retrieval and record identifiers, origins and missing fields.
  • Validate structures and document conversion losses.
  • Approve calculator, resource limits and execution separately.
  • Check energy compatibility or displacement-force completeness before analysis.
  • Report hull, equilibrium and phonon conclusions with their distinct assumptions.
  • Preserve manifests, numerical failures and unresolved decisions.
  • Reserve experimental claims for appropriate experimental evidence.

Continue with these related guides: AI for Computational Chemistry Workflows.

Sources

The Materials Project MCP README supports its tool inventory and deployment requirements. The ChatMOF README and SciAgents README support agent roles, dependencies and scope.

The named Pymatgen Skill, ASE Skill, Phonopy Skill and pycalphad Skill support their workflow contracts and scientific limitations.

Licence discrepancies can be inspected in the SciAgents licence, SciAgents package metadata, and ASE collection licence.