Skip to content

Getting Started with Agents

Install DataChain and its coding-agent skill to process data files and reuse saved results across sessions.

Installation

pip install datachain
uv add datachain

Install the skill for the current project, then start a new agent session:

datachain skill install --target claude --local
# Other targets: cursor, codex, copilot, pi

Omit --local to install for your user account. The default datachain skill includes recipes for inventories, media metadata, batched inference, video and recording readers, saved-result queries, updates and interrupted runs.

For shared team datasets and compute, the skill can also guide agents using Studio MCP. Studio is optional and currently invite-only; local workflows need no Studio account.

Your First Agent Task

Describe the task in your own words:

Inventory our deliveries across S3, GCS and Azure.
Report file counts and bytes by source, folder and extension without downloading contents.
Task DataChain pattern
One-off inventory list_storage() and one aggregation pass
Media metadata A typed file's get_info() inside map(), with prefetch=0
Batched inference A function with setup() and batch_map(..., batch=32)
Video and recordings Use VideoFile.get_frames or a Python MCAP reader in map/gen; retain source files and timestamps
Saved results Inspect datasets and recorded code, then query read_dataset()
File updates A compatible pipeline with read_storage(..., update=True, delta=True)
Interrupted runs Rerun the script to reuse checkpointed work

For questions about saved results or provenance, start from recorded datasets and code. Check external sources when current freshness is part of the request. Read the skill's installed SDK.md for API details and its SKILL.md for the model-change rebuild rule.

A small shell command is sufficient for questions it answers directly. Use DuckDB or pandas for ready-made tables.

Knowledge Base

Studio MCP provides shared dataset descriptions and knowledge search. Locally, agents inspect saved datasets with dc.datasets(), dc.describe() and dc.read_dataset().

Installing datachain replaces the retired core, knowledge and jobs skills in the selected target and scope, including their companion instructions. Existing datasets and dc-knowledge/ files are preserved.

Next Steps