LOCAL-FIRST RESEARCH WORKSPACE

Turn materials questions into
research you can trace.

Start with papers, formulations, and lab records. Let an agent use the right Skills and tools while keeping each source, action, and result in one research project.

Start a task Check release status

The desktop app is free, and 0.3.0 installers connect to the production platform. MX Points purchases and cloud models are currently limited to pilot accounts.

MaterialsX / Research workspace01 — 03
Actual MaterialsX new research task screen with stages and six starter tasks
Actual development UI · New research task
0198bundled scientific Skills
02176potential and resource listings
032selectable agent engines

START WITH A QUESTION

What research will you move forward today?

Start with a question, choose a stage, and add the materials you need before sending.

New here? Begin with one question

Choose a task, then add your material system, goal, or files in the composer. Edit before sending, and verify data and citations in the result.

Choose a research stage

Begin with a problem, papers, or existing data. No prompt-writing experience needed.

6 common tasks
Getting startedTopics & literatureMaterials dataExperiments & analysisStructures & simulationPapers & outputs
01

Clarify my research question

Turn an early idea into goals, constraints, and next steps

Describe the material, target property, and what you already have.
02

Find relevant papers

Set keywords, search scope, and trustworthy sources

Start with a topic; distinguish abstracts from full papers.
03

Read a paper with me

Understand methods, results, evidence, and limitations

Requires a PDF; return to source pages for key values.
04

Understand an experiment table

Check units, missing values, replicates, and outliers

Requires data; explain columns and test conditions first.
05

Design a first controlled experiment

Specify variables, controls, repeats, and measures

Add available equipment, time, cost, and materials.
06

Check a research claim

Separate evidence, inference, and missing validation

Provide the claim and source; identify a minimal test.

In the desktop app, choosing a card only fills the composer. It does not send a message or run tools. You can also describe your own task.

Actual MaterialsX new research task screen with stages and six starter tasks
Actual development UI · New research task
  1. 01Choose a task
  2. 02Add context or files
  3. 03Review and send
  4. 04Verify results and sources

WHY MATERIALSX

Research is a chain of decisions, not a single answer.

MaterialsX connects objectives, input data, tool runs, and conclusions. Unverified records keep their status, and computed results carry their applicability limits.

01

Start with evidence

Keep page numbers, original values, and sources close to the claims they support.

02

See the method

Review the plan, Skill, tool calls, and receipts for each step of the task.

03

Check the result

Keep charts, structures, and reports in the project, alongside gaps and scientific caveats.

A RESEARCH FLOW

Four clear steps from question to deliverable.

Handle simple questions directly. For complex work, define goals, constraints, and acceptance criteria first.

01

Define the goal

Name the material system, question, target measures, and limits.

02

Gather evidence

Choose papers, project files, or configured MOOS records.

03

Run a method

Use authorized Skills, analysis tools, or eligible models.

04

Review delivery

Inspect data, charts, structures, reports, and unresolved gaps.

CAPABILITIES

Built for real materials research tasks.

Bring common workflows together while making data, runtime, and human-review requirements explicit.

Papers & files

Extract with source locations

Work with PDF, DOCX, spreadsheets, and text. Extracted passages retain page, paragraph, or sheet locations for review.

PDF limit: 100 MiB / 2,000 pages
Research workflows

98 bundled Skills

Browse bilingual descriptions and examples by topic, invoke a Skill with @, or install one from a local folder or GitHub.

Some Skills need optional scientific dependencies
Research data

Connect MOOS records

A separate read-only MCP service can query experiments, formulations, processes, properties, images, and simulations.

MOOS service configuration required
Materials models

Screen before calculating

Explore 100 model and resource entries, then check domain, elements, license, weights, and runtime before choosing a route.

A catalog entry is not an installed checkpoint
Agent engines

Local or platform models

Choose Pi or the bundled Codex App Server. Connect LM Studio locally or a configured platform service.

Tool-use capability depends on the model
Structures & review

See both results and limits

Inspect atomic structures and 3D artifacts while checking units, evidence, convergence, and method applicability.

Computation needs matching runtimes and weights

MATERIALSX 0.3 / SYSTEM FLOW

From research question to MX Points settlement.

The website explains the product; the desktop handles research and purchases. An independent account gateway owns billing, LiteLLM connects model suppliers, and a separate console reconciles transactions.

01 / PUBLIC

Official website

Explains capabilities, release status, and downloads. It holds no wallet balances, payment credentials, or supplier keys.

Public presentation
02 / USER

MaterialsX desktop

One workspace for projects, sources, Skills, and MOOS. Signed-in users see their MX wallet, orders, eligible models, and usage.

Research and purchase
03 / CONTROL

Account and billing gateway

Authenticates users and payment callbacks, owns the MX ledger, locks prices and reserves points before model calls, then settles trusted usage.

Financial source of truth
04 / OPERATIONS

Separate billing console

Role-based staff views for orders, wallets, usage, refunds, prices, reconciliation, and audit. Writes go through the account gateway.

Finance and operations
05 / MODEL

LiteLLM and supply

An independent proxy routes to RootFlowAI. Upstream API keys stay in private server configuration; its admin UI is for authorized operators.

Routing and procurement

How one paid model task closes the loop

  1. 01

    Top up

    Choose ¥10 / 50 / 100 / 500 on desktop. A verified WeChat payment grants 100 / 500 / 1000 / 5000 MX Points once.

  2. 02

    Select a model

    The desktop shows eligible models. The gateway locks the route and procurement, FX, and retail price versions.

  3. 03

    Reserve and route

    The gateway reserves points, then LiteLLM calls the supplier with a server-side key. No upstream key reaches the client.

  4. 04

    Settle usage

    Trusted input, output, cache-read, and cache-create token counts determine the charge. Missing usage stays pending reconciliation.

  5. 05

    Reconcile

    The billing console compares consumed revenue, supplier invoices, and refunds, recording differences, audit events, and measurable gross profit.

Deployment and data boundaries

  • Static public site; local projects and models stay on user devices.
  • Account gateway, billing Web, and LiteLLM deploy, health-check, and roll back separately.
  • MX balances and orders live only in the MaterialsX business database; LiteLLM uses its own database.
  • Supplier keys remain in LiteLLM private configuration; keys and procurement costs never reach public or client pages.

Buy MX Points once; pay for each model’s usage

Fixed conversion: ¥1 = 10 MX Points. These are retail snapshot rates as of 2026-10-07, in MX Points per million tokens of each category: input / output / cache read / cache create. Actual eligibility is shown after sign-in.

Model and tierInput / output / cache read / cache createCurrent status
gpt-5.6-sol175 / 1050 / 17.5 / 218.75Limited access
gpt-5.6-sol ≥272k350 / 1575 / 35 / 437.5Limited access
claude-opus-5-5140 / 700 / 7 / 175Limited access
claude-fable-5-1350 / 1750 / 8.75 / 437.5Limited access

The approximate 50% of official USD reference prices is a historical comparison. Supplier quota uses the confirmed 1 quota = ¥1 assumption for cost estimates. Top-ups, earned revenue, and supplier costs are recorded separately; gross profit is not net profit.

The 0.3.0 installers connect to the production platform. A pilot account completed a ¥10 WeChat payment, received one MX Points grant, and settled a GPT 5.6 charge. Three models are available to pilot accounts at the approved rates. Actual supplier invoices, refunds, and cross-platform installation remain to be verified.

01 / 03

MACHINE-LEARNED POTENTIALS

Find a suitable atomic potential before you calculate.

Explore general and specialized potentials in the materials model catalog. Screen by system, elements, periodicity, and task. Inspect weights, license, and runtime status before downloading or loading an eligible model.

  1. 01

    Describe the structure and task: energy, forces, or relaxation

  2. 02

    Check elements, charge, spin, training domain, and license

  3. 03

    Download and load eligible weights for a bounded run

  4. 04

    Inspect energy, forces, convergence, report, and 3D structure

A catalog listing is not an installed checkpoint. Some entries have no public weights or cannot run on your device. AI recommendations still require applicability checks and real execution receipts.

Actual MaterialsX machine-learned potential catalog with model and installation status
Actual development UI · Potential catalog
02 / 03

RESEARCH DATA · MOOS MCP

Use past formulations and processes to plan the next experiment.

With MOOS MCP configured, query formulations, original amounts, preparation steps, test conditions, and source evidence in read-only mode. Compare compatible records before proposing constrained formulation candidates and validation experiments.

  1. 01

    Search relevant formulations and process records

  2. 02

    Keep units, test conditions, page citations, and review status

  3. 03

    Compare compatible samples and property differences

  4. 04

    Propose candidates and controls for experimental validation

MOOS is a separate service and must be configured. An optimization candidate is a research suggestion, not a validated formulation. Historical records are never silently changed.

Actual MaterialsX Data and MCP screen showing MOOS read-only connection status
Actual development UI · MOOS connection
03 / 03

LOCAL LLM

Run your language model on your own computer.

Load a model in LM Studio and start its local API. In MaterialsX, select Local model, enter the service URL, and test the connection. Research files and conversations can be processed on-device.

  1. 01

    Load a model suited to your task in LM Studio

  2. 02

    Start the local API, usually http://localhost:1234/v1

  3. 03

    Check model and agent-engine compatibility in MaterialsX

Chat-model weights are not bundled. Tool use, file reading, and long-task reliability depend on the model and runtime; run the compatibility check first.

Actual MaterialsX local model settings with URL and compatibility controls
Actual development UI · Local model settings

DOWNLOAD STATUS

Choose your installer.

The 0.3.0 installers for macOS Apple Silicon and Windows x64 connect to the production platform and are available on GitHub Releases. Signing, notarization, and cross-device acceptance remain open.

View 0.3.0 release notes Read the setup guide

macOS

Apple Silicon

0.3.0 · available

Check the installer architecture, SHA-256 and release notes before installing.

MaterialsX-0.3.0-mac-arm64.dmg1787.3 MB
Download DMG

Windows

x64

0.3.0 · available

Check the installer architecture, SHA-256 and release notes before installing.

MaterialsX-0.3.0-win-x64.exe740.6 MB
Download EXE

Linux

x64

No public installer yet

Check the release page for the exact version and architecture.

Check the downloaded file against the SHA-256 list: View SHA256SUMS.txt

No chat-model weights are bundled. Load a local model in LM Studio or another compatible tool. Scientific outputs require human review.

GET STARTED

Begin with a local project.

01

Choose a project folder

Create a research task so conversations, inputs, and artifacts stay in your workspace.

02

Connect a model

Load a model in LM Studio and start its local API, then check the connection in MaterialsX settings.

03

Ask a checkable question

Choose a Skill or state a goal directly, such as extracting experimental values with page citations.

Example prompt

@materials-literature-rpsme-json Extract the experimental formulations from this paper with original amounts and evidence page numbers.

MaterialsX

FAQ

A few things to know before you begin.

Is MaterialsX free to use?

The desktop app is free, and local models can run on your own computer. Eligible platform models use MX Points based on four token categories. Purchases are currently limited to the local test environment; the signed-in catalog shows actual eligibility and rates.

Are all 100 materials models installed and ready?

No. The 100 items are bilingual catalog entries for potentials, models, and related resources. Weights, dependencies, licenses, and execution eligibility must be checked individually.

Can I connect my own local model?

Yes. MaterialsX supports the local LM Studio API. Tool calling and long-task reliability depend on the model, protocol, and runtime; run a compatibility check first.

Is MOOS data available automatically?

No. The separate MOOS MCP service must be deployed and configured. Queries are read-only, and records awaiting review keep that status.

Does it replace scientific judgment?

No. MaterialsX helps organize evidence, run eligible methods, and preserve receipts. Researchers must still verify experiments, model applicability, and final scientific claims.