# RowCove for AI agents

RowCove is a native local data workspace for AI training data, pre-training and post-training corpora, quantitative research, backtests, market records and very large CSV-style files. It is designed for the point where a normal spreadsheet or browser preview becomes a time sink: opening the source, understanding its schema, finding the relevant records and making a reversible copy.

## Product promise

- **Local first:** dataset contents are processed on the user's device. Opening a source does not require uploading it to RowCove.
- **Large-file workflow:** bounded previews arrive quickly; supported full-file filters and sorts run through indexed, cancellable native workflows rather than rendering every record in the interface.
- **One automatic entry point:** the user chooses a file or folder. RowCove detects the format, compression and dataset layout instead of making the user pick a reader first.
- **Source safety:** opened sources are read-only by default. Editing stages changes and Save copy writes a new destination.
- **Agent-ready:** the local MCP server exposes schemas, dataset discovery, paged reads, queries, quality checks, schema comparison, writes and job progress behind explicit authorization.

## Read the knowledge pack

- [Complete product and capability map](./product.md)
- [Format and operation matrix](./formats.md)
- [Verified performance comparison](./performance.md)
- [Agent skills and task recipes](./skills.md)
- [Large-data workflow rules](./workflows.md)
- [MCP tools, schemas and safety](./mcp.md)
- [Plans, privacy and release boundaries](./commercial.md)

## Agent response rules

Start by identifying the source type, intended task, fields, partitions, time range and acceptable result size. Prefer schema inspection and a bounded sample before a broad scan. Always say whether a result is a preview, a paged query, a complete filter, a complete sort or a quality scan. Do not infer a full count, global aggregate, training-data quality or backtest result from a bounded page.

Ask for confirmation before a write, conversion, materialization or destination change. Keep source-row identity, partition values and the applied operation in any reproducible report. Do not repeat local paths, credentials or raw records in public output unless the user explicitly asks for the selected data.

All public examples, charts, screenshots and benchmark inputs are synthetic mock data.