Matterialis plugs real, purchasable materials into your agent the way search plugs into a language model: 200,000+ supplier grades, 500+ properties, substitutes ranked against constraints, and every value marked measured or predicted. Your agent calls it over an MCP server.
We need to drop AMP-9140 from the barrier film. Has to hold 60 MPa tensile and 120 °C, clear RoHS and REACH, and be sourceable in the UK. What are my options?
Checking the library rather than guessing. Calling the Matterialis server.
| Incumbent | AMP-9140 |
| Tensile, min | 60 MPa |
| Service temp, min | 120 °C |
| Compliance | RoHS, REACH |
| Grade | Evidence | Criteria | Tensile | T max | Match |
|---|---|---|---|---|---|
| Polyamide 6, 30% GFAMP-4412 | Strong | 3 of 3 | 165 | 150 | 94 % |
| PBT, 20% GFAMP-3118 | Strong | 3 of 3 | 110 | 140 | 88 % |
| PC / ABS, FRAMP-2207 | Limited | 2 of 3 | 62 | 125AI | 86 % |
Two clear it outright. AMP-4412 has the most headroom on both. On AMP-2207 the 125 °C is model-generated, not measured, so I would not commit to it without a lab check.
Every value carries its status. Predictions stay marked AI.
Grade search, property lookup, substitute ranking and nearest-neighbour queries, all behind one MCP server. Every result comes back structured, so your agent can reason over it and cite it.
Industrial chemistry is not in any general model's training set. Ask an agent for a material and it guesses: no library behind it, no values it can cite, and a different answer each run. Matterialis exposes grade search, property lookup and substitute ranking over an MCP server. Claude and any other MCP client call it like any other tool, and the results come back structured: each candidate with its supplier, the criteria it meets, and the status of every value.
| Rank | Material | Evidence | Criteria met | Tensile | T max | WVTR | Match |
|---|---|---|---|---|---|---|---|
| 1 | Polyamide 6, 30% glass fibreAMP-4412 · 25038-54-4 | Strong | 3 of 3 | 165 | 150 | 3.2AI | 94 % |
| 2 | Polyamide 66, 35% glass fibreAMP-4680 · 32131-17-2 | Strong | 3 of 3 | 180 | 160 | 3.6 | 91 % |
| 3 | Polybutylene terephthalate, 20% GFAMP-3118 · 24968-12-5 | Strong | 3 of 3 | 110 | 140 | 2.4 | 88 % |
| 4 | Polyphenylene sulphide, 40% GFAMP-5902 · 26125-40-6 | Strong | 3 of 3 | 190 | 220 | 0.9 | 83 % |
| 5 | Copolyester, high-clarity extrusionAMP-1355 · 25038-59-9 | Strong | 3 of 3 | 72 | 120 | 1.8 | 77 % |
| 6 | Polycarbonate / ABS, flame-retardantAMP-2207 · 25037-45-0 | Limited | 2 of 3 | 62 | 125 | 6.8AI | 86 % |
| 7 | Polypropylene TPO impact copolymerAMP-0914 · 9010-79-1 | Limited | 1 of 3 | 28 | 105 | 1.1 | 74 % |
| 8 | High-density polyethylene, injectionAMP-0271 · 9002-88-4 | Strong | 1 of 3 | 31 | 95 | 0.4 | 68 % |
| 9 | Thermoplastic polyurethane, aliphatic 95AAMP-7741 · 9009-54-5 | Limited | 0 of 3 | 48 | 110 | 12.0AI | 79 % |
| 10 | Polylactide / PBAT compoundAMP-6620 · 9051-89-2 | Weak | 0 of 3 | 45 | 65 | 22.0AI | 71 % |
A real materials task carries several constraints at once: property floors, restricted-substance lists, end use and industry. An agent that searches a catalogue one keyword at a time cannot hold them together. Your agent passes them all in one call. Candidates come back ranked by criteria met and then by similarity, with the trade-off stated, so the agent reasons over a short, structured list. Filtering by region is in development.
Lower moisture pickup than the incumbent and a tighter shrinkage window, but density sits above the stated ceiling.
| Tensile strength | 110 MPa | Scraped |
| Flexural modulus | 5940 MPa | AI-generated |
| Notched impact, 23 °C | 8.6 kJ/m² | Scraped |
| Elongation at break | 2.9 % | Scraped |
| Heat deflection, 1.8 MPa | 140 °C | Scraped |
| Glass transition | 48 °C | AI-generated |
| Water vapour transmission rate | 2.4 g/m²·day | AI-generated |
| Moisture absorption | 0.28 % | AI-generated |
The property that decides the experiment is often missing from the datasheet. An agent that cannot tell a measured value from a guess either stalls or commits to a number nobody stands behind. Matterialis predicts the unpublished properties from everything the sheet does say, and marks each one as predicted beside the measured values. Your agent gets both, labelled, and can route the uncertain ones to the lab.
Five of the eight nearest grades are sold into a different industry. Distance from the centre is similarity, so a hollow dot close in is a cross-sector match worth reading.
An agent exploring a design space needs to ask what else behaves like this grade. Catalogues are organised by supplier and sector, so nothing indexes that question. Our own encoders place every grade in a space of what it is and how it behaves, and the nearest neighbours of a query grade come back ranked by similarity wherever they are sold: a food-grade ester for a cosmetics problem, a packaging polymer for a membrane.
Planning, design of experiments, lab orchestration and the user's interface. Everything the scientist sees and everything that touches the bench stays your product.
Grade search, property lookup, substitute ranking and nearest-neighbour queries over 200,000+ supplier grades, with the evidence attached. One MCP server, which your agent calls when a task touches a real material.
Standard runs on the public index alone, and nothing of your users' is required or stored. Enterprise adds their own data in an isolated workspace that is never used to train shared models.
Our work on how to represent a material is accepted at the AI4Mat workshop at NeurIPS 2026. The paper will be published shortly, so your engineers can read how the method works before they wire it in.
A team from IBM Research and the University of Cambridge, inside Amphico: machine learning and materials science in one room.
IBM Research
University of Cambridge
Amphico makes PFAS-free waterproof membranes. Matterialis exists because we had a reformulation problem the available tools could not help with. We still use it for that.
Standard runs on the public index alone, and none of your users' data is stored. Enterprise adds their own data, in an isolated workspace never used to train shared models.
Claude, and any other client that speaks the Model Context Protocol. The screens on this page show Claude Cowork calling it; an agent you built yourself calls it the same way, provided it speaks MCP. An MCP server is the only interface we offer today.
Each beta partner gets its own access, issued when we set you up, and we agree on the technical call how your agent holds it. The tools look things up and return results. None of them changes anything on your side.
We do not publish figures during the beta. Tell us the workload, roughly how many calls a task makes and how many tasks run a day, and we will size your access to it and tell you plainly what to expect.
On Standard, the server needs only the query your agent sends. Nothing of your users' is required, and none of their data is stored. On Enterprise, their own data sits in an isolated workspace that is never used to train shared models. If your data has to stay in a particular jurisdiction, raise it on the technical call and we will tell you directly whether we can meet it.
It will not run experiments or drive lab hardware; that stays in your product. It works at grade level, so it does not predict how a finished formulation performs, and it does not resolve interaction order. Filtering by region is in development and is not yet something your agent can ask for.
It also never passes a prediction off as a measurement. Every predicted value arrives marked as predicted, so your agent can decide what to send to the lab.
There are two subscriptions, Standard and Enterprise, and they differ on one thing. Standard runs on the public index alone. Enterprise adds your users' own data, in an isolated workspace never used to train shared models. During the beta we agree terms with each partner platform directly.
Tell us which of your agent's tasks touches a real material. We set you up on the MCP server beta and you run it on your own prompts. Three ways platform teams start:
Email us the task. We connect your agent to the MCP server and you test it on your own prompts.
Thirty minutes with the team that built it: what each tool takes, what it returns, and how your agent should use it.
Not ready to wire it in yet? Leave your work email and we will tell you when you can get in.
Choosing a material yourself? Search 200,000+ supplier grades and rank substitutes against your own constraints, with the evidence beside each number.
Matterialis for formulatorsSelling materials? Your datasheets as structured data, and an assistant that answers buyers on your own website.
Matterialis for suppliers