Skip to content

Skill and MCP

The skill ships with pip install datachain and teaches agents to inspect saved results, process files and save new datasets. It guides local SDK workflows and, when connected, shared team workflows through Studio MCP. Local use needs no Studio account. Studio is optional and currently invite-only.

Local SDK and Studio MCP

Local SDK (OSS) MCP (Studio)
Ships with pip install datachain Studio
Reads Local .datachain/db (SQLite) Centralized data warehouse + shared Knowledge Base
Compute Local machine, parallel threads Jobs with or without a cluster; clusters scale across dozens or hundreds of CPU/GPU machines
Scale One project's SQLite registry Billions of typed records (LAION-scale corpora)
Access control Filesystem only Teams, namespaces, per-team ACLs
Audience Local data processing and saved results Teams; large datasets; headless agent services

The Skill (OSS)

pip install datachain
datachain skill install --target claude     # also: cursor, codex

The Skill is the agent-side instruction package. Installing it places files into the agent harness's directory (for Claude Code: ~/.claude/skills/datachain/). The default skill teaches the agent to inspect the Dataset DB, process data files, save expensive results and query them later.

Agents inspect saved datasets with dc.datasets(), dc.describe() and dc.read_dataset(). The skill does not maintain local markdown summaries.

Compute runs locally. map(), gen(), and agg() execute on the developer's machine through the in-process Compute Engine with parallel threads and async prefetch. Single-machine reach.

Local datasets remain available across sessions. For a shared team catalog, see Studio MCP below.

MCP (Studio)

Studio provides shared datasets and jobs through MCP, the Model Context Protocol supported by Claude Code, Cursor and Codex. The skill guides the agent to inspect shared results and relevant jobs before computing, then choose an available execution option. Shared results and job tracking work without a cluster; clusters are available for larger workloads. Generated knowledge descriptions help discover existing work; inspect dataset versions and producing code to verify it. See MCP Server for connection instructions for users with Studio access.

Studio adds four capabilities the OSS path does not have:

  • Distributed compute. Optional BYOC clusters run data processing across dozens or hundreds of CPU and GPU machines using the same DataChain code.
  • Billion-record registry. A full-scale data warehouse replaces SQLite. Filter, join, group_by, and similarity search run at warehouse speed over billions of typed records (LAION-scale corpora).
  • Teams and access control. Users belong to teams; datasets sit in namespaces; per-team ACLs determine who reads and writes what. The MCP surface enforces the same model as the Studio UI.
  • BYOC. Compute clusters and storage stay in your cloud account. Studio orchestrates without holding the data.

When You Need Each

Scenario Install
Solo developer Skill
Team needing access control across users and projects Skill + MCP
Billion-row datasets or distributed compute across many machines Skill + MCP
Headless production agent service MCP

See Also