Skip to content

One business.
One Twin.
Any AI.

Generic AI has never seen how your business works. MLX reads your systems and writes down what things mean — that record is the Twin. The AI you choose answers from it, nothing else — and the work it produces files back in as auditable records.

  1. 01Evidence arrivesYour systems connect — read-only. Data streams in from every edge.
  2. 02In suspensionIt is only data — mixed, ungoverned, not yet meaning.
  3. 03It organisesMLX proposes what things mean, layer by layer.
  4. 04Structure, heldThe rings are the Twin — what your business means, written down and held.
  5. 05Agents at workAsk in Sessions; hand agents tasks and automations that run on it.
  6. 06Filed back inAgent work returns to the Twin — records in Channels, Data Sets and Views, not threads.
  7. 07Every doorThe same Twin from ChatGPT, Claude, the CLI and MCP.

Front elevation — the AI Twin, live

Connected evidence sourcesYour source systems remain authoritative
Accounting · payments
Xero
SunSystems
Stripe
GoCardless
Documents · files
Microsoft 365
SharePoint
Google Drive
Knowledge · CRM
Google Docs
Google Sheets
Notion
HubSpot
Email · calendar
Gmail
Calendar
Outlook
Chat · meetings
Slack
Teams
Granola
Engineering · data
GitHub
Linear
Postgres
MySQL
SQL Server
Further sources scoped per activation

Mercury Labs has delivered for

  • HM Government logo
  • Department for Education logo
  • Department for Work and Pensions logo
  • Crown Commercial Service logo
  • IfATE logo
  • HVM Catapult logo
  • MTC logo
  • Harrison Parrott logo
  • PrimaryBid logo
  • Augmentive logo
  • Forestreet logo
  • Lucy Vail Floristry logo

Client work by Mercury Labs, the studio that builds MLX.

01 · How you start

Start with one useful part of your business.

Connect your priority systems, let MLX author and prove the Twin, then review the release and switch it on — two focused weeks.

  1. Connect priority systems

    Connect the systems that matter first, read-only. MLX maps what they can prove.

  2. MLX authors and proves the Twin

    MLX works out how your systems fit together, writes down what things mean, and tests every result against your own database. Anything it cannot prove is left out.

  3. Review the release, switch it on

    The sealed release stays dark until you have read it. Switch it on, and your approved AI tools answer from the Twin.

02 · Ask

Ask in a sentence. Read the whole working.

One question, one governed answer — the table, the totals and what they mean.

show aged receivables, ranked by value per customer

Aged receivables by customer

As at 11 August 2026 — GBP, ranked by total outstanding

RankCustomerInvoicesTotal ARCurrent1–30 days31–60 days
1Ridgeway University1£6,187.50£0.00£6,187.50£0.00
2DIISR – Small Business Services2£4,470.63£4,200.00£0.00£270.63
3Bayside Club2£3,434.00£3,434.00£0.00£0.00
4Port & Philip Freight2£2,245.00£0.00£2,245.00£0.00
5Boom FM1£1,623.75£1,623.75£0.00£0.00
6City Limousines4£1,169.95£650.00£0.00£269.95
7Basket Shop1£914.55£914.55£0.00£0.00
8Hamilton Smith Ltd1£550.00£550.00£0.00£0.00
9Rex Media Group1£550.00£550.00£0.00£0.00
10Marine Systems1£396.00£396.00£0.00£0.00
Total16£21,541.38£12,318.30£8,432.50£540.58
  • Overdue: £9,223.08, or 42.8% of total receivables.
  • Collection priority: Ridgeway University and Port & Philip Freight account for £8,432.50 of the overdue balance.
  • No 91+ day balances or missing due dates were found.

Basis: Xero sales invoices with a positive live AmountDue; voided and deleted invoices excluded. Aging uses invoice due dates. The synced invoice snapshot was last updated 9 August 2026 at approximately 06:53 UTC.

Retrieved with
MLX warehouse · Xero
Evaluate with TwinTraceBranch

03 · The receipt

Every answer shows its working.

The receipt is not a footnote — it is the architecture: which sources, which definition, which version, and when.

Specimen03 · The receipt · scale 2:1Governed answerissued by the Twin · as an agent receives itQuestionWhat was revenue last quarter?Answer£1,243,908 · Apr–Jun 2026Published definitionMonthly revenue · v2Data current to30 Jun 2026Evidence usedPublished business viewIf unsupportedSays so · never guessesEvidence referenceillustrative specimenMercury Labs · LondonAsked as writtenRule and versionWhere it came fromFigure with its periodFreshness statedNo plausible guessEVERY GOVERNED ANSWEREVIDENCE · ATTACHED01 · figures shown are examples, not customer data02 · an unsupported question is declared, never guessedMLX-DRW-003 · The receiptIssued byThe TwinProvenYour databaseStatusGoverned answer

Pan sideways to inspect at drawing scale.

Fig. 03Every governed answer carries its evidence.
Governed answerissued by the Twin · as an agent receives itQuestionWhat was revenue last quarter?Answer£1,243,908 · Apr–Jun 2026Published definitionMonthly revenue · v2Data current to30 Jun 2026Evidence usedPublished business viewIf unsupportedSays so · never guessesEvidence referenceillustrative specimenMercury Labs · London
Fig. 03Every governed answer carries its evidence.

04 · Behind the receipt

Nothing marks its own homework.

The receipt names a definition. This is who wrote it, proved it and judged it.

Exhibit — who builds the Twin, and who checks itSix workers · one chain
  1. 01Software measures

    It surveys your systems and tests every possible join across all your data. Sensitive values are fingerprinted first.

  2. 02One AI interprets

    It reads measurements, never your rows, and designs views around questions your team would ask.

  3. 03A compiler writes the code

    The exact views and tests are generated from that design — nothing improvised.

  4. 04Your database proves it

    Every view must run and reconcile on your own data.

  5. 05A second AI judges

    Separate from the author — it cannot repair or approve the work, only judge it.

  6. 06A gate admits

    Anything unproven is refused. What passes is sealed, and stays dark until you switch it on.

The reviewerNever the author

05 · In the work

Mention a View. The agent gets your records.

Type @, mention a View, and the agent works from exactly the records it shows — opened beside the conversation.

Renewal risk review

Across @Renewal command — what's at risk in the next 30 days?

Four renewals land in the next 30 days. Orbit Education and Alder & Co Logistics are already Renewed, and Brightwater Energy is on track at 86% for 22 September.

The one to watch is Northstar Gridworks — £84,000 renewing 14 September at 72%, health At risk. The next step on the record: resolve the DPA redlines.

Retrieved with
Run trace · 44 updates · 128219msHelp improve this answer

Ask about @Renewal command — or mention another View

Auto

Renewal command

TableBoardDashboard
TitleRenewal dateProbability

Renewal accounts · 12 records · scope pinned to this Session

06 · The feed

The discussion, where the records are.

Every channel has a feed, and the Twin is a member — it posts after overnight syncs and answers when mentioned, receipt attached.

Renewal Operations Demo

Data setsTrace dataChannel access

Share an update with the channel…

Maya Chen · Today, 09:12

Pushed the Northstar decision call to Thursday — legal wants one more pass on the DPA before we counter.

Sam Okafor · 09:31

@MLX if Northstar slips past quarter end, where does committed Q3 land?

MLX · 09:31

£84,000 would leave the quarter. Renewed so far stays at £79,000 — Orbit Education and Alder & Co Logistics — with Brightwater Energy's £96,000 still to land on 22 September.

Retrieved with

MLX · Today, 06:54

Refreshed aged receivables after this morning's Xero sync — £21,541.38 outstanding, 42.8% overdue. Collection priority is still Ridgeway University and Port & Philip Freight.

07 · Chaos to records

Chaos in. Records out.

What your systems scatter — and what your agents produce — the Twin files into tables you define.

  • XeroINV-0219 · Ridgeway University
  • GmailRE: Northstar renewal
  • SunSystemsJNL 2214 · £84,000.00
  • DocsRenewals playbook v7
  • Slack#renewals · 47 unread
  • HubSpotNorthstar Gridworks · deal
  • StripePayout · £12,304.55
  • NotionRenewal owners table

Scattered across your systems

The Twin organises

TitleAnnual valueStage

Filed as records you keep

08 · Keep

You always know what you are looking at.

Every number on screen says which records it came from and how they were filtered.

Renewal Operations Demo

Data setsTrace dataChannel access
Renewal commandRenewal accounts · 12 records
TitleHealthAnnual valueRenewal dateRenewal probabilityAccount leadNext stepStage

09 · Any model, every door

Any model. Any door. Same answer.

Change the model, or come through a different door — the answer cannot move, because it comes from the Twin, not the model's memory.

Governed business contextv2
  • MLXthe session
  • ChatGPTover MCP
  • Claudeover MCP
  • Geminiover MCP
  • CLItwin query
  • SDK · APIyour code

Every route reads the same governed context — and every answer carries the same permissions and audit trail, whichever door it came through.

10 · Where it lives

A server that is yours alone.

Every Twin runs on its own dedicated MLX server — in the region you choose, connected to your systems read-only, with an audit trail end to end. Nothing pooled, nothing shared.

Deployment — MLX serverUK · EU · US

Choose a hosting region

Server
Dedicated · yours alone
Region
UK region
Connections
Read-only

One region per deployment. The flag marks server geography only.

11 — The system

Five parts. One Twin.

Everything above is one system — a governed copy of your business that AI can safely use. Five parts carry it.

  1. 01

    Lives with your business

    Always on, on a server that is yours alone — UK, EU or US.

  2. 02

    Meaning, written down

    It proposes definitions in plain words; every version is kept and inspectable.

  3. 03

    Any AI, your choice

    OpenAI, Anthropic, Google or your own hardware — swap models, keep the meaning.

  4. 04

    Files records, not threads

    Agents must write into tables you define, so knowledge stays findable.

  5. 05

    In the tools you already use

    MLX, Claude, ChatGPT, terminal or code — same answers, same permissions.

12 · Questions

Asked, answered.

The questions regulated teams ask first — answered plainly.

Where does our data go?
MLX connects read-only to the systems you approve and keeps each target scoped to your organisation. Those source systems remain authoritative. When a workflow uses an external model provider, the context needed for that request follows the provider path configured for your deployment.
Do we need to migrate our legacy systems first?
No wholesale or rip-and-replace migration is required. A focused activation connects the priority systems you approve, starting read-only, and brings the relevant evidence into the Twin while those systems remain authoritative. Mercury Labs engineers then work with your experts to turn that evidence into governed business context — written down, tested and versioned.
Can we choose models and control their cost?
Yes. The governed business context stays separate from the model provider. You can use stronger models for demanding work, lower-cost models for routine tasks or evaluate another configured provider without rebuilding that context. Material changes are tested before rollout.
What does this cost?
MLX is a scoped engagement. Mercury Labs engineers support the activation and operate the configured target, while model usage is recorded where provider evidence is available. Talk to us for a proposal based on the systems, workflows and operating requirements in scope.
Why not just use ChatGPT or Copilot?
Use them where they fit. The Twin gives your AI tools the governed business context they normally lack, including agreed definitions, source freshness and evidence about what each answer used — and agents can file their work back into channels as auditable records. It complements the AI surface rather than replacing it.
We're regulated. Can we actually use this?
MLX starts scoped to your organisation and read-only, and business context ships as sealed releases you can review before they are activated. Normal governed queries do not include browser, terminal or write access. A regulated deployment still requires approval of the exact data and model paths, evidence retention and operating controls.
What do the AI agents actually see?
During the build, the authoring agents never read your rows. Sensitive values are converted to keyed fingerprints before anything is compared, and the agents work from descriptions and measurements — so relationships can be found without raw values reaching them. Once the Twin is live, agents work inside your organisation's permissions: they can pick the published, governed context or explore your connected records free-form.
Which systems can you connect?
Current connectors: Xero, SunSystems, Stripe, GoCardless, Microsoft 365 and SharePoint, Google Drive, Docs and Sheets, Gmail, Google Calendar, Outlook, Slack, Microsoft Teams, Granola, Notion, Linear and GitHub. Legacy ledgers such as Infor SunSystems stay queryable without migration; exact scope is confirmed before activation.
What happens when the Twin cannot support a question?
It says so. A governed answer identifies the published product, version and freshness it used. If that product does not support the requested question or evidence, the Twin returns 'not supported' instead of producing a plausible guess. Agents can still explore the underlying records free-form — that work is simply not labelled a governed answer.
Two ways to start

Build the Twin your AI can rely on.

Activation

A focused two-week activation starts with one priority workflow and the systems behind it.

Partnership

Partner enquiries are handled by Mercury Labs, the London-based AI engineering studio that builds and deploys MLX.