Answers about customers and payments live in systems your AI assistant cannot see.

Ask your AI assistant about your own business.

Connect Claude, or any AI assistant that speaks MCP, and ask about your customers, payments and balances in plain language. STRAX answers from its synced company model through structured, read-only queries, and each key sees only what its data policy allows. Every question is written to the audit trail, and access can be switched off at any time.

Request a demo → All features →
A question answered within the key's data policy A finance analyst asks Claude which merchants settled late this week. STRAX applies the data policy of the analyst's key: two allowed entities, a hidden bank-account column and a row rule for one region taken from the key. It runs a structured read-only query with a row cap and a timeout and returns two merchants with a note on how fresh the data is. Each call is written to the audit trail, and the rows go to the model provider behind the AI client. Asked in Claude, with the key finance-gauteng: "Which merchants settled late this week?" This key's data policy entities: Merchant, Settlement hidden: bank account number rows: region = Gauteng, from the key Structured query aggregate Settlement by Merchant filter: settled late, this week read-only · row cap · timeout Audit trail 10:42 describe · Settlement 10:42 aggregate · read-only 10:42 2 rows · ok 2 merchants settled late Merchant 1043 · 3 days late Merchant 2210 · 2 days late synced 4 minutes ago The rows go to the model provider behind the AI client. sensitive data: use the built-in assistant with a local model A plain question, checked against the key's policy. A hidden column, and rows limited by the key. A structured query that can only read. The answer, its freshness, and where rows go.

Questions in plain language

Which merchants settled late this week? What is this customer's balance, and when did they last pay? The assistant finds the right business entity in STRAX, looks at how it is described, and then queries, aggregates or searches it. Answers come from the hub's synced copy of your systems, with a note on how fresh the data is, so you know whether a figure is from this morning or last month.

  • Customers, merchants, payments and balances are found, described, queried, aggregated and searched as business entities.
  • Every answer carries a freshness note from the last synchronisation.

Each key sees only what you allow

Access is set per key with a data policy. You choose the entities a key may read, hide the columns it must never see, and filter rows by attributes on the key itself, such as a region or a merchant. If a row rule needs an attribute the key does not carry, STRAX refuses the call.

  • A list of allowed entities for each key.
  • Hidden columns never appear in an answer.
  • Row rules filled from the key's attributes, for example a finance key limited to one region.
Live Each key sees only what you allow Entities · hidden columns · row rules
One data policy applied to two keys A data policy allows two entities, hides the bank account column and filters rows by the region carried on the key. The key finance-gauteng gets two Gauteng merchants and no bank account column. The key finance-ops carries no region, so STRAX refuses its call and returns no rows. Data policy set per key in the STRAX console may read: Merchant, Settlement hidden: bank account number rows: region comes from the key the same policy on two finance keys finance-gauteng · region: Gauteng MerchantRegionSettled Merchant 1043Gauteng3 days late Merchant 2210Gauteng2 days late bank accounthidden, never returned finance-ops · no region on the key asks for late settlements by merchant refused: the row rule needs a region no rows returned, and the refusal is recorded Each key reads under its own data policy. Rows are filtered by the key's own region. A key without that attribute is refused.
Allowed entities, a hidden column and a row rule filled from the key: the key without a region gets nothing.

Structured and read-only

The assistant asks structured, parameterised questions, and STRAX runs them under a read-only login with row caps and timeouts. There is no free-form database access by default. A confused prompt can return a poor answer, but it cannot change a record, and a heavy question stops at its cap or its timeout.

  • Parameterised queries under a read-only login.
  • Row caps and timeouts on every call.
Live Structured and read-only Parameterised · read-only login · caps and timeouts
Questions run as structured, read-only queries A confused request asks the assistant to mark a merchant as settled. It becomes a parameterised query that runs under a read-only login, so the status is read and nothing is written. A heavy question for every payment since 2019 runs under the same rules and stops at its row cap. A confused request "Mark Merchant 1043 as settled." asked in Claude, key finance-gauteng Structured query query Settlement by merchant parameter: merchant 1043 Read-only login status read, nothing written no record changed ✓ A heavy question "List every payment since 2019." same key, same rules Structured query query Payment · from 2019 row cap · timeout Stopped at the row cap capped, then stopped caps and timeouts apply to every call Questions become structured, parameterised queries. A read-only login cannot change a record. Heavy questions stop at a row cap or timeout.
A confused request cannot change a record, and a heavy question stops at its row cap or its timeout.

Reports as PDFs

Ask for a report your team already runs in STRAX and the assistant has it rendered on request. The PDF comes back as a single-use link that expires shortly afterwards, so the report is never left at an open address.

  • Uses the saved reports your team already maintains, with their parameters.
  • Each link works once and expires shortly afterwards.
Live Reports as PDFs Saved report · rendered on request · single-use link
A saved report delivered as a single-use PDF link Asked in Claude for August's settlement report for Gauteng, the assistant runs the monthly settlements report your finance team already maintains, with its region and month parameters. STRAX renders a six-page PDF and returns a single-use link. The link opens once, and a later attempt finds it expired. The question "August's settlement report for Gauteng." asked in Claude, key finance-gauteng Saved report Monthly settlements parameters: Gauteng, August maintained by your finance team Rendered as PDF Monthly settlements · 6 pages rendered on request The link 10:14 · link 7f3a… returned to Claude works once 10:15 · opened · PDF downloaded 10:40 · opened again link expired Ask for a report your team already runs. It is rendered to PDF on request. The link works once, then it expires.
A saved report rendered on request, returned as a link that works once and then expires.

Know where your data goes

The rows an AI assistant reads over MCP go to the model provider behind that assistant, the same as anything else you type into it. Choose a provider whose data terms you are comfortable with, and keep sensitive entities and columns off the keys you give to hosted clients. For data that must stay on your servers, use the assistant built into STRAX with a local model.

  • Over MCP, the rows an answer uses reach the client's model provider under that provider's terms.
  • Data policies decide what reaches the client at all.
  • The built-in STRAX assistant can run on a local model, so its questions and answers stay on your servers.

The built-in assistant, with a local model →

Live Know where your data goes Over MCP · built-in assistant on a local model
Where the rows go, over MCP and with a local model Inside your servers, STRAX holds the synced company model. Over MCP, the key's data policy decides which rows leave, and those rows reach Claude and the model provider behind it, under that provider's terms. The built-in STRAX assistant can run on a local model, so its questions and answers stay on your servers. Sensitive columns are best kept off keys given to hosted clients. your servers Your STRAX installation synced company model Merchant · Settlement · Customer The key's data policy decides which rows leave Claude, over MCP 2 rows in the answer outside your servers The model provider under its data terms like anything typed into the client Built-in assistant same question, asked in STRAX Local model runs on your own servers questions and answers stay on your servers ✓ Choosing per key hosted clients: keep sensitive columns off the key data that must stay: the built-in assistant The key's data policy decides what leaves at all. Rows read over MCP reach the client's model provider. With a local model, answers stay on your servers.
Rows read over MCP reach the client's model provider, while the built-in assistant on a local model keeps them on your servers.

Audited and switchable

Every question, including the ones STRAX refuses, is written to the tamper-evident audit trail under the key's name. Rate limits apply per key, and an administrator can switch MCP access off for every client from the next request. Data access over MCP is licensed separately from development access.

  • Refused calls are recorded as well as answered ones.
  • Rate limits per key, and one switch that stops every client.

Request a quote for MCP data access →

Keep exploring

More in Run

Know what is happening and fix it early.

The monitoring screens show records moving, jobs running, queues building and errors as they happen, and when a job fails the assistant proposes the fix.

Operations & governance →

Answers about the company, with the sources shown.

AI Insight reads the strategy, processes, workflows, systems and current data, answers with its sources shown, and lists every area a change would touch.

AI Insight →

Describe the application and the assistant builds it.

Tell the assistant what the business needs and it plans, builds, tests and promotes the tables, workflows, portal and reports, with you approving each step.

AI Assistant →

See it on your own systems.

A demo takes about an hour. Bring the failure that last woke someone up and we will show you how STRAX handles it.

Request a demo → Request a quote