SDC Agents SMB

Purpose-scoped ADK agents for SDC4 data operations — SMB Edition.

Designed for personal and small/medium business usage. Uses a local LLM via Ollama instead of a Google API key, while connecting to the commercial SDCStudio SaaS backend for catalog, validation, and assembly APIs.

Positioning

Agents

8 purpose-scoped agents with 39+ tools (core + ToolsetHub plugins):

Introspect dynamically loads ToolsetHub plugins — install [notion], [sheets], or [airtable] for SMB-native datasource support.

Features

Core Pipeline

• Introspect datasources with 13-field standardized column analysis and 10 type inference patterns
• Discover matching catalog components with type compatibility scoring
• HITL review gate for billable minting operations — see costs before committing
• Assemble data models via the SDCStudio Assembly API (sync + async with hybrid polling)
• Download published data model packages (.zip with XSD, XML, JSON, JSON-LD, HTML, SHA1)
• Generate XML instances from mapped datasource records
• Validate instances against XSD 1.1 schemas via VaaS API (deterministic, not probabilistic)
• Distribute artifact packages to Fuseki, Neo4j, REST APIs, or filesystem

SMB-Native Datasources (ToolsetHub)

• Notion — database properties, relations, rollups, select options (pip install sdc-agents-smb[notion])
• Google Sheets — headers, inferred column types, sheet metadata (pip install sdc-agents-smb[sheets])
• Airtable — field types, linked records, formula/lookup fields (pip install sdc-agents-smb[airtable])
• Community extensible — add HubSpot, QuickBooks, Salesforce by following the reference pattern

Automation

• Scheduler — cron-based pipeline automation via APScheduler (sdc-agents schedule run)
• Notifications — push status to Slack webhooks, Telegram bots, or SMTP email
• Pipeline templates — 7 bundled workflows (sdc-agents pipeline run healthcare-csv -p datasource=patients)

Data Governance

• Schema drift detection — compare current structure against cached previous introspection, alerts on changes
• Data annotations — agents auto-detect anomalies (null violations, mixed date formats, sentinel values); users add manual notes; annotations persist across sessions and auto-merge into future introspections
• Cross-datasource lineage — track data flow from source through mapping, generation, validation, to distribution
• Compliance reports — generate JSON/Markdown/HTML evidence from audit + lineage logs (sdc-agents compliance report)
• Append-only audit — every tool call logged with credential redaction to .sdc-cache/audit.jsonl

Integrations

• MCP server mode — serve any agent as an MCP server for Claude Desktop, Cursor, etc.
• Audit dashboard — web UI for browsing, filtering, and exporting audit records (sdc-agents audit serve)
• OpenClaw skill — 9-tool bridge exposing SDC tools to OpenClaw's messaging platform ecosystem

Quick Start

1. Install Ollama and pull a model

# Install Ollama: https://ollama.com/download
ollama pull gemma4:26b

2. Install SDC Agents SMB

pip install sdc-agents-smb

# Optional extras:
pip install sdc-agents-smb[knowledge]                    # PDF, DOCX, ChromaDB
pip install sdc-agents-smb[notion,sheets,airtable]       # SMB datasources
pip install sdc-agents-smb[dashboard]                    # Audit dashboard web UI

3. Configure

cp sdc-agents.example.yaml sdc-agents.yaml
# Edit sdc-agents.yaml with your SDCStudio URL, API key, and datasources

4. Run

# MCP mode — serve an agent as an MCP server
sdc-agents serve --mcp catalog
sdc-agents serve --mcp introspect

# Check configuration and installed toolsets
sdc-agents info
sdc-agents toolset list
sdc-agents validate-config

# Run a pipeline template
sdc-agents pipeline list
sdc-agents pipeline run healthcare-csv -p datasource=patient_csv

# Start the scheduler
sdc-agents schedule list
sdc-agents schedule run

# View audit log and dashboard
sdc-agents audit show --last 24h
sdc-agents audit serve --port 8080

# Manage data annotations
sdc-agents annotate list-all
sdc-agents annotate add my_csv email "EU rows use comma decimal separator"

# Assembly review workflow
sdc-agents assembly list-pending
sdc-agents assembly review quarterly_model
sdc-agents assembly approve quarterly_model

# Generate compliance report
sdc-agents compliance report --format html --last 30d -o report.html

5. ADK mode (standalone agent)

from sdc_agents.agents.catalog import create_catalog_agent
from sdc_agents.common.config import load_config

config = load_config("sdc-agents.yaml")
agent = create_catalog_agent(config)
# model defaults to ollama_chat/gemma4:26b from config

Model Configuration

The default model is ollama_chat/gemma4:26b. Configure in sdc-agents.yaml:

model:
  default: "ollama_chat/gemma4:26b"
  ollama_base_url: "http://localhost:11434"

Tested Models

Any Ollama model with tool-calling support should work. Use the ollama_chat/ prefix for chat models.

Security Model

• Purpose scoping — each agent has a narrow tool set, no mega-agent
• Security isolation — no agent has both datasource access AND network access
• Private project enforcement — created components go to non-public SDCStudio projects only
• Read-only datasources — SQL write operations rejected; all introspection is read-only
• Credential redaction — audit logger redacts connection, token, key, password, secret
• Path confinement — validation/distribution restricted to configured output directory
• Append-only audit — every tool call logged to .sdc-cache/audit.jsonl
• ToolsetHub scope enforcement — plugins declare network hosts, datasource types, and file access; violations rejected at load time

Development

Environment

This repo needs its own environment, and that is not a preference. SDC_Agents and SDC_AgentsSMB both provide the module sdc_agents, so they cannot coexist: install both and import sdc_agents resolves to whichever landed last, which means a test run can silently exercise the wrong repo while reporting success. It also depends on google-adk, which SDCStudio removed deliberately, so the SDCStudio environment cannot run these tests either.

conda env create -f environment.yml
conda activate SDC_AgentsSMB
pip install -e ".[dev]"
pytest -q

CI runs ruff check, then black --check, then pytest. Run the formatters before pushing.

Releasing

Two steps, and merging alone ships nothing. release.yml triggers on a tag, not on a push to main. Merging publishes the Docker image and nothing else, which looks like a finished release.

# after the PR is merged and main carries the new version
git checkout main && git pull
git tag -a v4.0.1 -m "4.0.1"   # must match `version` in pyproject.toml
git push origin v4.0.1

The tag builds and publishes to PyPI. __version__ derives from the installed distribution, so only pyproject.toml needs the bump.

Documentation

• Repository Guide — catalog of all SDC FOSS repos
• Architecture Overview — how the stack fits together
• ClawFeatures — competitive positioning vs OpenClaw
• PRD — product requirements document

License

Apache License 2.0 — see LICENSE.