Maincode (Matilda) Review: Australia's Homegrown AI Assistant — Promising, but Still Early

Verdict (TL;DR): Maincode is building something genuinely interesting: an AI assistant trained and hosted in Australia, built from first principles by an Australian team, with real commercial traction — including Australia's first authenticated agentic payment transaction via Mastercard. But at the time of this review Matilda was in private beta, pricing was not published, general public access was not available, and the company had walked back its earlier "sovereign AI" claims. Matilda entered open beta on 14 July 2026 — see Update history, 22 September 2026. As an idea and an infrastructure play, Maincode deserves serious attention. As a product you can deploy today and compare to ChatGPT or Claude, it cannot yet make that case. Watch this space.

Score: 14/25 — ⭐⭐⭐ Good

Last reviewed: 26 June 2026 by Adam.

What Maincode is

Maincode is a Melbourne-based applied AI research company building Matilda, an AI assistant and foundational language model trained in Australia. The company was founded by Dave Lemphers — a former Microsoft and AWS data scientist — and is majority-owned by Ed Craven, co-founder of online casino platform Stake.com.

Maincode announced Matilda in August 2025, initially describing it as "Australia's first built from scratch, totally sovereign foundational AI model." The company has since walked back the "sovereign" framing — Lemphers told Information Age in October 2025 that Maincode had "moved away from the language of sovereignty" due to partnerships with international hardware, connectivity, and software providers. The company's current positioning is "Australian-made AI": built in Australia, trained in Australia, by an Australian team.

Matilda is described as a foundational AI assistant with deep contextual understanding, multi-step reasoning, and the ability to take actions through agentic frameworks. Maincode's business model appears to be a combination of its own consumer/business-facing AI assistant and custom model development for enterprise clients.

The company's infrastructure sits in a Telstra data centre in Melbourne, powered by AMD Instinct GPU hardware — a deployment Maincode calls the "MC-2 AI Factory," involving a reported AU$30 million investment.

What it costs in Australia

There is no published pricing for Matilda. At the time of this review the product was in private beta with a waitlist. That changed on 14 July 2026, when Maincode opened Matilda’s public beta — see Update history, 22 September 2026. The availability statements in this review describe June 2026 and are now out of date.

Maincode's website describes Matilda as an AI assistant with "early access coming soon." The beta application form is open but access is by invitation. There is no free tier, no published subscription price, and no indication of when general availability will arrive.

For enterprise customers seeking custom model development, pricing is presumably negotiated directly with the Maincode team. (no pricing information is available from primary sources. Contact Maincode at hello@maincode.com for commercial enquiries.)

How it works in practice

Based on Maincode's own materials, the beta SXSW Sydney demonstration (October 2025), and the Mastercard deployment, here is what Matilda appears to offer:

As an AI assistant: Matilda is designed for deep contextual understanding across long conversations, multi-step reasoning with inspectable steps, and learning user workflows over time. The company emphasises that Matilda "asks before it acts" — a deliberate design choice around agentic safety. The interface appears designed to be approachable without prompt engineering.

As a foundational model for custom deployments: This appears to be Maincode's primary commercial traction point. Matilda has been used by Mastercard Australia as the underlying model for Australia's first authenticated agentic payment transactions, completed in January 2026. In those transactions, Matilda-powered agents booked cinema tickets and accommodation using CBA and Westpac payment cards through Mastercard's Agent Pay framework. Lemphers has also mentioned pilots with "a very large Australian bank" and "a very large Australian technology and software company."

Infrastructure: Built on AMD Instinct MI355 accelerators and AMD EPYC 9675F CPUs, hosted in Telstra's Melbourne data centre. Maincode describes sub-200ms response latency globally from its inference cluster. All GPU infrastructure is in Australia.

What is less clear from primary sources is the model's benchmark performance relative to global frontier models, the training data composition in detail (Lemphers confirmed at SXSW that Matilda uses global data including Project Gutenberg — not only Australian data), and the production capabilities of the consumer-facing assistant beyond the beta environment.

What it does well

The infrastructure is genuinely Australian. GPU cluster in a Telstra Melbourne data centre. Training pipeline in Australia. The team is Australian. Whatever the caveats around "sovereignty," the physical infrastructure and operational team are onshore in a way that matters for data residency and national capability arguments.

Real commercial deployment already. The Mastercard Agent Pay transaction in January 2026 is not a demo or a press release — it is a live, authenticated agentic payment using an Australian LLM as the underlying model. That is meaningful real-world validation at a scale that matters. Maincode being selected as the AI layer for Australia's first agentic commerce deployment alongside Commonwealth Bank of Australia and Westpac is a credible commercial signal.

The agentic design thinking is sound. Maincode's stated design principles — "asks before it acts, checks when unsure" — address the biggest risk with agentic AI: unprompted action with real consequences. Building safety-conscious agentic behaviour into the model from the start is the right architectural choice.

First-mover position in Australian LLM development. There are very few organisations in Australia that have trained a foundational language model from scratch. Maincode is one of them. The research residency programme and university collaborations (Adelaide, UNSW, Victorian state government) suggest a serious research operation, not just a marketing exercise.

The "Australian-made" positioning is honest. After initially over-claiming "totally sovereign," Maincode has landed on a more defensible position: made by Australians, trained in Australia, hosted in Australia, by an Australian-owned company. That is accurate and meaningful, even if the training data is not exclusively Australian.

What it doesn't do well

You can use it now — and until 14 July 2026 you could not. This was the defining limitation of this review, and it has expired. Matilda opened its public beta on 14 July 2026: matilda.maincode.com takes an open sign-up with no invitation, maincode.com offers a desktop download and an API, and the waitlist at maincode.com/subscribe is gone (404). What has not changed is the money: there is still no published pricing and no /pricing page. (Corrected 22 Sep 2026 — this paragraph said “You cannot use it yet… they are on a waitlist” and called that the fundamental difference between Maincode and every other assistant on this site.)

Capability relative to global models is unknown. Maincode has not published benchmark results comparing Matilda to GPT-4o, Claude Sonnet, Gemini Pro, or any other mainstream frontier model. The SXSW demo involved an AFL-data-trained model, which is illustrative but not a full capability assessment. Until Maincode publishes independent benchmarks or makes the model available for evaluation, capability comparisons are impossible to make honestly. (no independent capability benchmarks found at time of review.)

The "sovereign AI" walk-back raised questions it has not fully answered. Maincode initially marketed Matilda as trained on "sovereign Australian data." Lemphers later confirmed this was not accurate — the training data includes global sources. The website text was not updated for some time after this clarification. For organisations making procurement decisions based on data provenance, the original messaging and subsequent clarification matter. Maincode should publish a clear, primary-source description of exactly what data Matilda was trained on.

Funding source requires disclosure. Maincode is majority-owned by Ed Craven, co-founder of Stake.com, an online gambling platform. This is publicly reported and not a hidden fact. But it is a relevant piece of context for government agencies or regulated organisations evaluating Maincode as a supplier, and for any reader considering whether the company's backing affects its long-term strategic priorities.

No pricing transparency. Not unusual for an early-stage company, but pricing opacity combined with beta-only access makes any cost-benefit assessment impossible. Buyers cannot begin a business case.

Australian fit

The Australian fit case for Maincode is real but requires careful reading. The GPU cluster is in Australia. The training was done in Australia. The team is Australian. The company is registered and operating in Melbourne. Processing happens onshore.

Where it gets more complicated: the training data is not exclusively Australian, the "sovereign" framing has been retired, and the company uses international hardware and software partners (AMD, Google Cloud infrastructure). None of these things make Maincode less Australian, but they do mean the original framing — "totally sovereign, end-to-end Australian AI" — does not survive scrutiny.

The honest version is: Maincode is an Australian company building an Australian AI assistant, with Australian infrastructure and an Australian team. The data that trained it includes global sources. The hardware partners are international. The company is open to international commercial partnerships and exports.

For buyers who need a model where no data touches overseas infrastructure or overseas partners at any layer, Maincode's current architecture should be investigated in detail before assuming it meets that requirement.

For buyers who want an Australian company, Australian-hosted infrastructure, and Australian-operated AI — and understand that the training data is not exclusively domestic — Maincode is the most credible local option in the AI assistant category.

Australian Fit score: 4/5 — Infrastructure onshore, team Australian, commercial partnerships building. Deducted one point for the training data ambiguity and the unresolved gap between the original sovereignty marketing and the clarified reality.

Alternatives

Until Matilda reaches general availability, Australian individuals and businesses evaluating AI assistants will be comparing it against the global front-runners:

ChatGPT — OpenAI's AI assistant. The most widely used AI tool globally. No Australian data residency. AUD pricing available. The benchmark every new entrant is compared against.

Claude — Anthropic's AI assistant. Strong reasoning and long-context performance. Anthropic has established a Sydney presence. No Australian data residency by default. AUD pricing available.

Microsoft Copilot — Microsoft's AI assistant, integrated across the Microsoft 365 ecosystem. Microsoft has Australian datacentre regions and offers data residency options for enterprise customers in Australia. The most practical choice for organisations already in the Microsoft stack.

Grok — xAI's AI assistant, included with X Premium. No Australian data residency. Relevant as a comparison for general-purpose AI assistant use.

Bottom line

Maincode deserves credit for doing something genuinely hard: building a foundational language model from scratch in Australia, standing up GPU infrastructure onshore, and landing a real commercial deployment with Mastercard before even reaching public availability. That is a real achievement, not a startup press release.

The honest assessment in June 2026 was that Matilda was not yet a product an Australian business or individual could evaluate, compare or buy. That stopped being true on 14 July 2026, when the public beta opened. It can now be evaluated and compared; it still cannot be priced, because Maincode publishes no pricing and no public benchmarks. The "sovereign AI" story has been partially retracted. The funding source is relevant context.

Maincode is a company to watch closely — the infrastructure play, the agentic focus, and the commercial partnerships suggest a serious business, not a marketing exercise. If Matilda reaches general availability with competitive capability benchmarks and clear pricing, this review score changes materially.

For now: hold. Check back in six months.

Recommended for: Enterprise organisations interested in custom Australian-trained model development for specific use cases; developers or researchers wanting early access to a local foundational model for experimentation.

Not yet recommended for: General business or personal AI assistant use — not because you cannot get it, which you now can, but because nobody here has yet tested it and Maincode publishes no pricing. For everyday needs use ChatGPT, Claude or Microsoft Copilot until this review has been redone against the live product.

Score

Note on scoring: the low marks on Value for Money and Ease of Use reflected product unavailability, not poor design, and this review said they should be reassessed at general availability. General availability arrived on 14 July 2026, so that reassessment is now owed — it requires using the product, which is a re-review rather than a correction, and the total of 14/25 should be read as provisional until then.

Frequently asked questions

Is Maincode's Matilda available to use in Australia?

Not any more. Matilda was in private beta when this review was written; its public beta opened on 14 July 2026. Australians can sign up directly at matilda.maincode.com, download a desktop app, or use the API — no waitlist and no invitation. The old waitlist page, maincode.com/subscribe, now returns a 404. Pricing is still unpublished.

Is Maincode's Matilda really sovereign Australian AI?

Maincode has stepped back from the 'sovereign AI' label. CEO Dave Lemphers confirmed in October 2025 that Matilda was not trained exclusively on Australian data — training data includes global sources such as Project Gutenberg. The GPU infrastructure is housed in a Telstra Melbourne data centre, and the team is Australian. Maincode's current positioning is 'Australian-made AI' — built and trained in Australia by an Australian company — rather than fully sovereign.

Has Maincode been used in any real commercial deployments?

Yes. In January 2026, Mastercard completed Australia's first authenticated agent-led payment transactions using Maincode's Matilda as the underlying AI model, through Mastercard's Agent Pay program. Transactions involved Commonwealth Bank and Westpac payment cards for cinema tickets and accommodation bookings. Maincode is also reported to be working on pilot models for a major Australian bank and a major Australian technology company.

Sources

• Maincode — Official website and Matilda description: https://maincode.com/ (primary) — Wayback 11 July 2026
• Maincode — beta availability status: the cited maincode.com/subscribe returns HTTP 404 as of 22 September 2026. Superseded by Maincode’s own announcement, “Introducing Matilda’s Open Beta: Australian AI, Built for Trust”, 14 July 2026 (primary), read live 22 September 2026: “Matilda is available at matilda.maincode.com and on the Apple App Store.” matilda.maincode.com resolves to a live sign-in, checked the same day. Last capture of the dead waitlist page: Wayback 7 November 2025 — context only, captured seven months before this review’s check date
• Information Age (ACS) — "Maincode debuts Matilda AI after ditching 'sovereign' label," Tom Williams, 15 Oct 2025: https://ia.acs.org.au/article/2025/maincode-debuts-matilda-ai-after-ditching--sovereign--label.html (secondary — CEO statements, SXSW demo, data sovereignty clarification) — Wayback 15 May 2026
• SmartCompany — "Aussie LLM Maincode powers Mastercard's agent-led payments trial": https://www.smartcompany.com.au/artificial-intelligence/mastercard-agent-pay-ai-maincode/ (secondary — Mastercard deployment) — Wayback 1 July 2026
• Mastercard Newsroom — "Mastercard accelerates AI-powered commerce with Australia's first authenticated agentic transactions," Jan 2026: https://www.mastercard.com/news/ap/en/newsroom/press-releases/en/2026/mastercard-accelerates-ai-powered-commerce-with-australia-s-first-authenticated-agentic-transactions-using-agent-pay/ (primary — Mastercard Agent Pay Australia)
• Data Center Dynamics — "Maincode to deploy AMD hardware at Telstra data center in Melbourne": https://www.datacenterdynamics.com/en/news/australian-ai-startup-maincode-to-deploy-amd-hardware-at-telstra-data-center-in-melbourne/ (secondary — infrastructure details)
• AMD Case Study — "Maincode Builds An AI Factory for Australia with AMD": https://www.amd.com/en/resources/case-studies/maincode-mc-2-ai-factory.html (primary/secondary — MC-2 hardware specs) — Wayback 10 July 2026
• Startup Daily — "Australia's large language model, Matilda, joins the global AI race": https://www.startupdaily.net/topic/artificial-intelligence-machine-learning/australias-large-language-model-matilda-joins-the-global-ai-race-as-maincodes-dave-lemphers-outlines-his-national-ambitions/ (secondary — company background)
• Business News Australia — "Australian AI Matilda aims to reduce sovereign risk": https://www.businessnewsaustralia.com/articles/matilda-llm-challenges-chatgpt.html (secondary — company background, Ed Craven ownership) — Wayback 3 September 2026
• Maincode — AMD case study and research: https://maincode.com/ (primary — principles, research areas, Melbourne HQ) — Wayback 11 July 2026

No AUD pricing available to cite. No independent capability benchmarks available at time of review.

The AI Directory Australia is an independent editorial publication. We don't accept payment for positive reviews. Read our Methodology and Affiliate Disclosure.