Moorhold docs

Moorhold turns a stack of open-source data tools into one governed workspace for people and their AI agents. You declare each data source once and Moorhold wires it into every tool. AI agents propose changes, and a person approves them. It runs on infrastructure you own, and you can take it out again.

Moorhold is in early access. The latest signed release is 2026.11.4, and organisations can install it themselves with an access key. Moorhold is free during early access, and we'll tell you well before that changes.

Moorhold is built for problems like these:

  • Lock-in. A platform that runs only on a vendor's cloud, at the vendor's prices, is hard to leave.
  • Assembly work. Putting open-source tools together into one platform is real work, even with no licence fees.
  • A login per tool. Every tool has its own login and its own settings.
  • The same connection, typed again. One database connection is typed into four places.
  • No line from pipeline to dashboard. Nothing connects a pipeline to the table it fills or the dashboard that reads it.

Moorhold is the layer between your people and the tools. It doesn't replace the query engine, the scheduler, the dashboards or the catalogue. It never forks or rebuilds a tool. It signs people in, wires each tool to the others through their own interfaces, keeps them in the state you declared, and records what changed and why.

People sign in through Keycloak, and every tool in the stack uses that sign-in. Groups are defined once and projected into each tool's own roles. Each tool keeps its own permissions: Moorhold doesn't add a new permission model on top. A launchpad shows each tool, whether it's healthy, and opens it in its own interface.

You declare a data source once, as plain configuration: where it is, which file on the server holds its password, and which tools use it. Moorhold reads an API source's password from a file:// reference under /etc/moorhold/secrets, a folder only root can read. sudo moorhold apply gives each password only to the service that loads that source, and Moorhold's declarations hold only the reference. Before anything changes, you see the plan, tool by tool. A reconciler compares what you declared with what each tool really has, and reports any difference it finds. When the source goes, Moorhold removes what it created.

Sources can be REST APIs with tokens, paging and rate limits, loaded with dlt. The portal calls this kind of source a web service. An admin adds one in the portal, and someone with root runs sudo moorhold apply once to give it its password. A database source can't get its password on a server installed with Moorhold's installer.

AI agents are users of the workspace, with the same rules as the person they act for. An agent always acts for a named person, with that person's permissions and no more.

  • Agents connect through the Model Context Protocol (MCP), to each tool's own server where one exists.
  • Changes from agents are proposals. A person in the admin group approves them, and never the person who asked.
  • Tools that write are off by default.
  • Secrets are references only. A proposal with a secret's value in it is rejected.
  • Moorhold doesn't ship a model or an assistant. You bring your own agent and your own model endpoint, and you choose where its model runs.

Pipelines, tables and dashboards are connected in DataHub without anyone drawing the lines by hand. The tools report what they read and wrote, using the OpenLineage standard. The audit log brings together the portal's own actions, sign-in and token events, and the query engine's record of what was read. Every agent action is attributed to the person it acted for.

You should be able to stop using Moorhold and keep everything it set up. You can export all of the portal's state as plain, declarative configuration, and that export is tested. Moorhold is designed so the stack keeps working if you remove it, because the tools are unmodified and configured through their own interfaces. Removing Moorhold from a running stack hasn't been tested.

Everything runs on servers you control. Moorhold runs on a single virtual machine with Docker Compose. Kubernetes isn't supported.

  • Secrets stay in the store you already use, and Moorhold keeps references, not values.
  • Each tool has its own service account, with no shared credentials.
  • Moorhold checks for updates only if an admin turns that on, or presses Check for updates. Your server asks get.moorhold.dev for the latest version number. It sends no data, but our site sees your server's address.

Moorhold itself is proprietary software from Allotment Technology Ltd. The tools it runs are established open-source projects, used unmodified, and you can use them without Moorhold.

Tool What it does
Keycloak Sign-in (OpenID Connect)
Trino Query engine
Apache Airflow Scheduling and pipelines
Apache Superset Dashboards and charts
DataHub Catalogue and lineage
PostgreSQL Database
dlt Loading data from APIs and databases
dbt Core Data models (SQL transformations)

These tools are under the Apache-2.0 licence, apart from PostgreSQL, which has its own PostgreSQL Licence. The stack also includes smaller supporting components under other licences. Each release records every component's licence and is licence-checked.

Each Moorhold release pins every tool to an exact version and is signed, so a tool you upgrade yourself leaves Moorhold's release line.

  • It isn't a hosted service. Moorhold runs on a server you control, at a cloud provider you choose or on your own hardware. Your data stays there.
  • It isn't open source. Moorhold is proprietary software built on open-source components.
  • It isn't finished. Not every feature described here is complete. We've built and tested them against copies of the stack on our own test servers, and none has run in a customer's environment.
  • It isn't a Windows app. There's no Windows app, and we aren't giving a date for one. The desktop app is for Macs with Apple silicon, and you can download it from moorhold.dev/download.