Enables natural-language querying of Databricks warehouses, grounding answers in schema and business definitions from the catalog.
Enables natural-language querying of PostgreSQL databases, grounding answers in schema and business definitions from the catalog.
Used for provisioning real Azure resources in production, part of the deployment stack.
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@data-agent-service Which support team resolves tickets fastest per our catalog definition?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Ask your data warehouse a question in English. Get an answer you can defend.
Natural-language questions over the warehouses, databases and APIs you already have — Fabric and Azure SQL, PostgreSQL, Databricks, and any REST service including a retrieval one — grounded in the glossary, metrics and schema held in OpenMetadata. Each query is authorized twice: role rules in the service, then the source itself under the caller's own identity. Any MCP client, unchanged against real Azure.
📖 Documentation site — the full
reference, also browsable as Markdown in docs/.
"Which support team resolves tickets fastest?" has a right answer only if everyone agrees what resolves means. Point a general-purpose SQL agent at the warehouse and it will infer that from column names, fluently and with no warning when it is wrong.
On this repository's own seeded data, that inference names the wrong team. Wall-clock elapsed time says Frontline is fastest and Billing is worst. The business's actual definition — Resolution Time, which excludes hours spent waiting on the customer — reverses it: Billing is fastest. A wrong winner is not a rounding error, and nothing in the answer would tell you it happened.
This service is built so that class of error is structurally hard rather than merely unlikely.
What you get | How | Proof |
Meaning comes from your catalog, not the model | Glossary terms, metric formulas and column descriptions are read from OpenMetadata at query time. Business semantics are never baked into a prompt |
|
Every answer runs as the person asking | The user's token is exchanged on-behalf-of all the way to the engine, so row and column permissions are the engine's decision, not the agent's |
|
It cannot write, wander, or work around a refusal | One read-only |
|
It answers with its reasoning attached | The figure, the definition applied, the tables it came from, and any caveat the catalog raised |
|
Any MCP client, no custom code | Claude, Cursor, VS Code and the SDKs connect over standard MCP with OAuth discovery | |
Runs on your laptop, deploys to real Azure unchanged | The whole stack runs on the emulator family; switching to Fabric, APIM and Entra is configuration, not a code path | |
Nothing here is claimed without something that checks it | Every capability carries a command that proves it; where something is designed but not built, the docs say so | 88 end-to-end witnesses, in CI on every push |
Status: Phases 0-16 landed — see docs/00-plan.md.
Related MCP server: MCP-DB-Server
make doctor # toolchain, docker, ~14 GB memory
make up # entra, keyvault, arm, fabric (+ SQL Server), OpenMetadata 1.13.2, apim
make status # "stack OK" is the verdictThen make seed, make test, make eval, make load, make ask Q="…" —
or make stack to do the whole bring-up from nothing, which is what CI runs.
Path | Purpose |
| Pinned, published images only — dependencies are used as-is |
| Every |
| Architecture, decisions, phases, evaluation, load, authz, extension |
| Quickstart, architecture, authorization, evaluation, load, MCP clients, adding a source, production, CI |
| The warehouse-query executor (Python and Go), and the contract both answer to |
| The agent, the accuracy suite, and the witnesses |
| Datasets, warehouse provisioning, OpenMetadata semantics, identity setup |
| Terraform for real Azure; |
| Four jobs; |
| The docs site — Astro + Starlight, generated from |
|
|
Emulators and OpenMetadata are never modified; suspected bugs go to docs/upstream-issues.md.
No emulator-only code paths. Standard protocols only (OIDC/OAuth2 incl. OBO, managed-identity App Service protocol, TDS FedAuth, ARM, Graph, OM REST/MCP). ENV=prod swaps .env and nothing else.
Built on entra-emulator, azure-keyvault-emulator, arm-emulator, fabric-emulator, azure-apim-emulator; composed per azure-emulators. Tier: leaf.
Apache-2.0.
This server cannot be installed
Unclaimed servers have limited discoverability.
If you are the server author, to access and configure the admin panel.
Query PostgreSQL databases in plain English — LLM-generated, safety-validated SQL.
The grounded data layer for any LLM: governed SQL, metrics, lineage and catalog over your data.
GibsonAI MCP server: manage your databases with natural language
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/calvinchengx/data-agent-service'
If you have feedback or need assistance with the MCP directory API, please join our Discord server