Meaning lives in the data layer.
DefinitionsYou write each definition once, and AI assistants, BI tools, APIs and applications all read it from the same place.
Meshly Data Stack
Meshly Data Stack gives them what a good colleague gives a new hire on day one.
They learn where your data lives, what your numbers mean and what they are allowed to see. Then they answer with the right figure and show the evidence behind it.
Four layers, from top to bottom. Meshly Data Stack is highlighted.
01 · Day one
You have found your desk, your laptop works and you have just poured your first coffee. Then the CFO stops by and asks you why the gross margin dropped in the third quarter.
Where do you even start? You do not know which systems hold the sales and cost data, or which of the dozens of tables with “revenue” in the name is the right one. You do not know whether margin in this company is counted before or after freight and discounts. You are not even sure whether you are allowed to open the customer-level figures.
A sensible new hire would walk over and ask a colleague. An AI agent has nobody to ask, so it makes a guess and presents that guess with complete confidence.
Without a semantic layer, every agent lives in this confusion each time you start it up.
02 · What a new hire needs
Your agents are already clever enough. What they lack is the knowledge that your organisation has built up over the years, and the platform can hand that knowledge over on every request.
A map. The new hire needs to know where the data lives and which tables can be trusted. An agent without a map invents table and field names that only look plausible.
A dictionary. The new hire needs to know what gross margin means in your company and which formula the finance team has agreed on. An agent without a dictionary picks a definition of its own.
A rulebook. The new hire needs to know which data they may use and for what purpose. An agent needs the same rules, and the platform enforces them on every request.
A calculator. The new hire would never estimate a quarterly figure in their head. An agent should not do that either, so a query engine computes every number from live data.
When an agent gets all four on every request, it stops guessing. We call this grounding, and we put it in the data layer so that every agent inherits it.
03 · The same question, with the layer in place
The CFO asks an AI assistant why gross margin dropped in the third quarter.
The assistant calls Meshly through the MCP gateway, and every call carries two credentials. One belongs to the CFO who asked the question, and the other belongs to the agent itself. Together they decide which data the answer may use and whether each field arrives in full, masked or empty.
The business glossary tells the agent that gross margin means net sales minus the cost of goods sold, and it points to the approved tables that implement the formula.
The policy engine weighs the person, the agent, the purpose and the classification of the data, and it records the decision with both identities.
The query engine calculates the figures from live data and applies the row filters and masking that the decision requires inside the engine.
The answer reaches the CFO together with the definition, the source tables and the lineage, so anyone can check how the figure was produced.
The model takes care of the language, and the platform takes care of the facts.
04 · The semantic layer
All of this lives in the data catalogue, and the pipelines keep it current as part of their normal work. Nobody has to maintain a separate wiki.
What is this? Every table and column has a plain-language description, so the agent finds the right data on its first try.
What does it mean? Each KPI has one approved definition in the business glossary, so the agent uses the same formula as your reports.
Where did it come from? Column-level lineage traces every figure from the source system all the way to the report.
Can I trust it? Quality tests and data contracts warn the agent when an input fails a check, before anyone acts on the result.
Who may see it? Sensitivity tags mark personal and confidential data, and the platform masks or withholds that data whenever the person or the agent lacks the right to see it.
Who owns it? Every dataset has a named owner, so questions and exceptions reach the right person.
One level deeper
The catalog, the policy engine, single sign-on and the MCP gateway, layer by layer.
Open ArchitectureThe pipelines, the dashboards and the agents all read the same metadata. Your organisation keeps one version of what its data means, and every tool uses it.
05 · How it fits
It helps to look at the platform from three heights, with each one a step further out.
Agents come in two kinds. Developers build programmed agents to run multi-step processes with fixed roles, approvals and business rules. Business teams assemble configurable agents from approved tools to handle everyday tasks. Both kinds reach the data through the same interface, follow the same policies and leave the same record, so an agent built in an afternoon follows exactly the same rules as one that took months to engineer.
06 · Built to stay open
A data platform usually outlives the project that bought it, so we designed this one to keep your options open for years.
You write each definition once, and AI assistants, BI tools, APIs and applications all read it from the same place.
Federated queries reach your existing sources, and you decide case by case whether a dataset is worth consolidating.
Every major query engine reads and writes Apache Iceberg, so the choices you make today remain reversible.
A dashboard, an API and an agent all get the same answer when they ask whether they may see something.
Meshly Data Stack has one annual price, so your bill stays the same when agents work around the clock.
The platform is built for Red Hat OpenShift, and Meshly is a Red Hat partner. It runs on your own servers, in your private cloud or in a fully airgapped network, and your keys and logs stay with you.
07 · Risk and compliance
Compliance remains your responsibility. The platform supports it with controls that are always on and with evidence that builds up as a by-product of normal operation.
| Regulation | How the platform supports it |
|---|---|
| GDPR | How the platform supports it: Personal data stays in your environment, sensitivity tags drive the masking, and the access log shows which person and which agent saw which personal data. |
| DORA | How the platform supports it: The platform runs in your own environment, it is built from components you can audit, and it keeps your data in open formats, which makes your exit plan credible. |
| NIS2 | How the platform supports it: Access is decided at one point in the query engine, and identities come from your own directory. You can audit the components yourself, and a software bill of materials for the release you run is available on request for your supplier assessment. |
| EU AI Act | How the platform supports it: Every answer carries its definition and lineage, and every agent action leaves a record. |
| Data Act | How the platform supports it: Your data, metadata and definitions stay in open formats that you can take anywhere. |
| Sector rules | How the platform supports it: Quality history and column-level lineage show where every reported figure came from. This supports the data-quality expectations in Solvency II and in banking supervision. |
Meshly provides the technical controls and the evidence. Your compliance also depends on your own environment and processes, and the Trust Center describes our own security practices in detail.
08 · In use
A financial services company moved from a metered cloud data warehouse to Meshly Data Stack in its own environment, and it now pays one flat annual price.
Agents assemble an insurer’s own risk and solvency assessment from the insurer’s own data. Actuaries review and approve the result, and every step and every data access is on record.
09 · Back to day one
This time the agent knows where the data lives and uses the margin formula your finance team agreed on. It sees only what both the CFO and the agent itself are allowed to see, and it lets the query engine compute the figures. The answer arrives with its evidence, so anyone can check it.
Your AI assistant is the right front door, and it needs to know what your data means to answer well. Meshly puts that meaning, the access rules and the arithmetic in the data layer, inside your own environment. Every agent you build inherits the layer, whichever assistant, framework or code you use, and your data stays in open formats that remain yours.
Configuration
The configuration decides which services run, and the hardware sizing is a separate choice that follows what you will actually do with the stack. We settle both with you in the first conversation.
Getting started
We deliver Meshly Data Stack to customers, so there is no self-service download. We plan it with you so that it fits the environment it has to live in.
Tell us about your sources, your users and where the platform has to run, and we will show you the stack on a live system.
Together we choose the services, the hardware sizing and the target, which can be Red Hat OpenShift, Kubernetes or your own servers.
We deploy with your team, connect the first sources and hand over the dashboard. Support and updates continue after go-live.
General notes
Licence, team and contracts live in one place. Every customer gets the Meshly Customer & Partner Portal, where they manage the licence, the people on the team, the contracts and support.
Release updates arrive in the portal. We publish new versions and their release notes there, so the team that runs the stack sees them first.
Partners onboard their own customers. A partner runs its customers in a tenant of its own. Ask about partnering
One level deeper
The technical pages describe what runs in the stack, how it connects to your sources and how your team operates it day to day.
Open the technical pagesThe Trust Center describes our security practices.
Request
Request
Tell us about a question that currently takes days to answer. We will map it onto your own data and show you how an agent answers it with the layer in place. We reply within two working days.