Introducing Agent Access in MyDataWork: a way to let an AI agent read the context of your data work — what you’re working on, how it connects, and what it’s for — without ever exposing your data.
The hardest part of putting an AI agent to work on real analytical problems isn’t the model. It’s context. An agent can be capable and still useless if it doesn’t understand the work it’s stepping into — which dashboards matter, what feeds what, which use case an asset serves, what’s fragile, what’s stalled.
That context already exists in MyDataWork. Every use case, asset, and lineage relationship you’ve organized is a map of your analytical work. Agent Access lets you hand that map — and only the map — to an AI agent, so it can reason about your work the way a well-briefed colleague would.
The key word is map. Agent Access exposes metadata only: the names, types, structure, lineage, and outcomes of your work. Never your data. Never your file contents. Never a single value inside a spreadsheet or a row from a warehouse table. The agent learns where things are and how they connect — never what’s inside them.
Agent Access lives in Setup. The header states the boundary up front — metadata only, never your data — and the four steps walk you from creating a key to connecting your agent.
What an agent actually receives
When you give an agent access, it can read:
- Your use cases — their objectives, status, and the outcomes they’re tracking
- The assets linked to each — file names, types, and tool (SQL, Python, Power BI, Tableau, Alteryx, and so on)
- The lineage between them — which work feeds which, inferred across tools
- Where your sources live — the warehouse tables and identifiers your work references
What it never receives: data values, file contents, query results, or rows. And — by design — the people. Stakeholder names and contact details are withheld from what an agent can read.
This isn’t a new trust posture for MyDataWork. The product’s own AI features have always worked this way: use-case text and asset metadata go to a model; file contents never do. Agent Access simply extends that same metadata-only boundary to your agent, over the Model Context Protocol (MCP), an open standard.
The preview shows exactly what an agent receives: your use cases, the assets linked to each, and how they connect — metadata only, no data values.
You decide what each agent sees — nothing by default
Agent Access is built around permission. Nothing is exposed until you choose to expose it.
You create a key for one specific agent. That key unlocks only the use cases you check — nothing else. You can create a different key for a different agent, each scoped to exactly what that agent should see. Every grant and every agent read is recorded in your audit log, and you can revoke or delete a key at any time. And before you connect anything, you can preview precisely what the agent will receive — the same view the agent gets, so there are no surprises.
Each key is created for one specific agent, in Live or Export mode. Keys can be revoked or deleted at any time.
You check exactly which use cases a key can read. Nothing is exposed until you choose it — and the key unlocks only what you’ve checked.
Two ways to connect: live, or export
There are two modes, and they suit different agents.
Live — your agent connects to MyDataWork over MCP and reads your current context each time it asks. Nothing to refresh: change a use case or link a new asset, and the agent sees it on its next read. This is the path for agents that connect with a credential — Claude Code, your team’s own agent, or an agent framework you’ve built on. Agent reads and exports draw on their own daily allowance — they never consume the AI credits you use for the chatbot or Leverage.
Export — you download your context as a point-in-time file and hand it to your agent: paste or upload it into the agent’s knowledge. No connection required. This is the simplest path, and it works with any AI tool that can read an attached file — including consumer chat apps like Claude Desktop and ChatGPT. Re-export whenever you want to refresh what the agent knows.
In Live mode, your agent connects to MyDataWork over MCP using the endpoint and your key. This works with bearer-key clients like Claude Code; Claude Desktop uses Export mode instead.
Seeing it work — with zero setup
You don’t have to connect anything to see what this does. With demo data loaded, MyDataWork includes a built-in illustration: the same question, put to the same agent, once without your context and once with it — so you can see exactly what the context changes.
Ask an agent to assess a forecast: “We’re going into the S&OP cycle. Prepare a readiness brief on our demand-forecast work: is it in good shape, what does it depend on, and what should I be cautious about relying on?”
Without context, an agent with only warehouse access can describe the tables — they’re present, recently refreshed, no schema anomalies. It can tell you the data is there. But it can’t tell you how the forecast is actually produced, what depends on it, or how the effort is tracking. It answers a narrower question than you asked.
With your MyDataWork context, the same agent traces the real chain of work: which assets feed the forecast, that a single workflow sits at the center with everything downstream depending on it, that the accuracy-improvement effort behind it is well short of its target, and that the chain references a source not yet in your catalog. It moves from “the data is there” to “here’s the state of the work, and here’s where I wouldn’t fully trust it yet.”
And it does that having read zero data values — only better context.
The built-in example: the same question put to the same agent, once without your context and once with it. With context, the agent traces the real forecast chain, the realization gap, and the uncatalogued dependency — having read zero data values.
Try it yourself
Anyone can reproduce this in a few minutes, on a free Explorer plan:
- Start a workspace and load the demo data (Setup → Load demo data).
- Open Setup → Agent access and read the built-in before/after example — no connection needed.
- To go further, create an export-mode key, choose the forecast use cases, and download your context file.
- Hand that file to your agent — upload it into a chat AI like Claude Desktop or ChatGPT — and ask it the forecast question above.
You’ll see an agent reason about your analytical work from context alone — and never touch your data.
On the free Explorer plan, Agent Access runs in Export mode: download a context file and hand it to any agent. Live connection unlocks on paid plans.
Where this fits
MyDataWork isn’t the agent, and it isn’t where your data lives. It’s the work-context layer — the map of your analytical work that an agent reads to orient itself, before it goes and does its job in the systems that hold the actual data. The agent reasons with MyDataWork, executes in your warehouse, and respects governance from wherever you govern it.
That’s the role: not the compute, not the data, not the agent — the context that makes all three smarter about your work. And it’s yours to expose, deliberately, one use case at a time, as metadata only — the map, never the keys.
Agent Access is available now on all MyDataWork plans. Export mode is included on the free Explorer plan; live connection unlocks on Solo and Team. [Start free →]