Projects

Every project declares its kind and apiVersion, plus a name, version, and status. The id is a unique identifier created for you during initialization. A description and a createdAt timestamp are optional.

A project belongs to one environment, and allowCrossEnvironment controls whether a job may combine sources and targets from different environments. maxTargetConnections caps concurrent connections to the target. driftPolicy declares what should happen when a source schema changes: a separate action for a patch, minor, or major change, with overrides per source, and per object or column where a single table needs different handling. Changes are classified and surfaced today; applying the policy automatically at run time is on the roadmap.

Sources refer to a data source such as a database or a file (API sources are on the roadmap). Amfelt uses data contracts to formally define the structure, ownership, and data quality expectations of data sources. Connection details live in the data contract or in the source’s own server, which takes precedence. Credentials stay out of both: secretRef points to them in a vault.

Targets are the data platforms you synchronize into, such as PostgreSQL or DuckDB. File-based destinations, such as Parquet files in an object store, are on the roadmap. A target names the plugin that writes to it, its server connection details, a secretRef for credentials, and an optional maxConnections cap.

Jobs move data between a named source and target. Each job selects objects from the source contract with a glob, names the data platform layer it populates (staging, history, bridge, business, or all). A job can record a writeMode, but no run acts on it yet: staging is reloaded in full every time. historyMode decides what the history layer keeps, and unlike writeMode it can be overridden per object (see Write Modes and History ). The target connector builds the history and bridge layers for each job, and the bridge layer may use business-friendly names to aid downstream data consumers. Scheduling belongs to the job: its runs are the strategy-and-schedule pairs that execute the selection (full ships today; incremental and cdc are on the roadmap), and maxParallel bounds how many objects it processes at once. A column can be excluded per object with columns.extract, so a sync does not select, stage, or historize it, or create it in the target. Primary-key, cursor, and event-time columns cannot be excluded, and an advise run still profiles every column of the tables it discovers. That is exclusion rather than protection. Amfelt does not mask, encrypt, or tokenize the values it does load; those are on the roadmap.

Three optional blocks generate files into the project repository: docs for Markdown documentation, dbt for a dbt project, and sql for the generated SQL. All three are off by default, and each lets you choose its directory.

Plugins declare which connector implementations the project uses and at what version, grouped by the role they play: source, target, and transformer. Selecting plugins per category is what makes the platform composable. The transformer role and the top-level transformers collection are Reserved : the format accepts them, but no transformer plugin ships today.