Quickstart¶
A database you already have → a governed answer with its provenance, in under an hour. Everything you need is public: the chart, the CLI, and this page.
Skeleton — commands not yet verified
This page is structure, not instructions. The steps, their order, and their verification points are fixed; the commands land in SV-P4·S6, drafted against the umbrella chart and run end-to-end on a real cluster first.
What you will have at the end¶
An agent answering a question about your data, and — the part that matters — the answer arriving with the record of every rule applied to produce it: the row-level filters, the masked columns, the caps. Governed, deterministic, auditable, and visible rather than asserted.
flowchart LR
DB[(Your database)] -->|1-3 import-schema| M[Your model in git]
M -->|4 serve| V[Veles]
V -->|5 MCP doors| A[Your agent]
A -->|6 ask| ANS[Answer + provenance]1. Prereqs¶
You need:
- A Kubernetes cluster. kind is fine — this whole quickstart runs locally.
- A reachable MSSQL or PostgreSQL database with any schema. Your own is the point; a toy one teaches you nothing about your data.
- An OIDC-capable identity provider. The chart bundles a dev Keycloak, which is acceptable here and nowhere near production.
You should now see…
TODO(S6) — each tool reporting a version at or above the floor, and
kubectl cluster-info naming your cluster.
2. Install¶
Install the umbrella chart from GHCR. One chart, the whole roster.
You should now see…
TODO(S6) — every pod Ready, and /ready green on the front door.
3. Import your schema¶
Point ttr import-schema at your database. You get back a db mirror (a
faithful, deterministic reflection of what is actually there), an er first cut
(the importer's proposal about how your tables relate), and a review
checklist naming every judgement call it made.
The checklist is the point. The importer never silently decides what your data means — it shows its work, grades its evidence, and asks.
Walk the checklist in VS Code (the extension links each item to the line it came from), then commit the model to a git repository. It is yours now.
You should now see…
TODO(S6) — a db layer mirroring your tables, an er layer proposing
relations with an evidence grade on each, and a checklist you can actually
read.
4. Serve the model¶
Point Veles at your model repository.
You should now see…
TODO(S6) — the catalog answering: get_model via ttr-meta-mcp returns
your model, or the Designer viewer draws it.
5. Connect an agent¶
Register the MCP doors in any MCP client, and forward your user's bearer token per the identity contract.
That forwarding is not a formality: the platform answers as your user, which is what makes the row filters in step 6 real rather than decorative.
You should now see…
TODO(S6) — the client listing the doors, and a whoami-style call coming
back as your user rather than as a service account.
6. Ask¶
Ask a natural question about your own data.
You should now see…
TODO(S6) — your answer, and its provenance attachment: the
pipelineWarnings trail naming every row filter, masked column, and cap the
platform applied on the way out.
This is the whole thesis, visible: the answer is not a guess about your data, it is a governed query over your semantics — and you can read exactly what was done to it.
7. Where next¶
You have the promise. Pick the job you actually have:
- Model — make the
erfirst cut mean what your business means. This is where the value compounds. - Connect — build a real agent: the full MCP surface, the identity contract, and the conformance suite as your test harness.
- Operate — run it for real: the values contract, your OIDC, policy in git, and one-question-one-trace.