Integrations

Datus Integrations

Datus is built on a plugin-first architecture. Beyond the dedicated databases and models pages, this page catalogues every other layer you can plug in — storage, embeddings, semantic layer, BI copilot, MCP protocol, skills, and observability.

Storage

Storage Backends — vector + relational

Dual-track storage keeps embeddings and relational metadata side-by-side. Swap the backend as your deployment grows.

Built-in

LanceDB + SQLite

Typedefault

PackageBuilt-in

NotesZero-config vector + relational store for local dev and single-node deployments.

PostgreSQL (pgvector)

Typepostgresql

Packagedatus-storage-postgresql

NotesProduction tier with per-namespace schema isolation and connection pooling.

Milvus

Typemilvus

PackagePlugin (v0.2.6)

NotesPurpose-built vector store for large-scale semantic retrieval.

Embeddings

Embedding Providers for context recall

Vectorize schemas, docs and metrics for semantic search. Mix cloud embeddings with local models to balance quality, cost and privacy.

OpenAI Embeddings

Modeltext-embedding-3-small / large

Dim1536 / 3072

Default

sentence-transformers

Modelall-MiniLM-L6-v2 (~100MB)

Dim384

Multilingual E5

Modelintfloat/multilingual-e5-large-instruct

Dim1024

BGE (zh / en)

ModelBAAI/bge-large-{zh,en}-v1.5

Dim1024

agent.ymlyaml

storage:
  # Database metadata embedding (cloud)
  database:
    registry_name: openai
    model_name: text-embedding-3-small
    dim_size: 1536

  # Document embedding (local, multilingual)
  document:
    model_name: intfloat/multilingual-e5-large-instruct
    dim_size: 1024

  # Metric embedding (local, fast)
  metric:
    model_name: all-MiniLM-L6-v2
    dim_size: 384

Semantic layer

Semantic Layer adapters

Bring your metric definitions into the agent's context. MetricFlow ships today; more adapters can be registered through Python entry points.

Ready

MetricFlow

Packagedatus-semantic-metricflow

Installpip install datus-semantic-metricflow

NotesMetricFlow-compatible YAML; joins into subject trees.

Core interface

list_metrics(path, limit, offset)— Enumerate metrics available in the semantic project.
get_dimensions(metric_name, path)— List every dimension a metric can slice by.
query_metrics(metrics, dimensions, ...)— Run metric queries with filters, time range and where clauses.
validate_semantic()— Validate the semantic configuration end-to-end.

BI copilot

BI Platform Copilot

Point Datus at an Apache Superset dashboard and it extracts every chart's SQL, builds a semantic model and emits two ready-to-use subagents — one for query, one for attribution analysis.

Superset only (today)GenSQL subagentGenReport + root-cause

Tableau, PowerBI and Looker adapters are on the roadmap.

terminalbash

# Deploy Superset + Postgres
helm upgrade --install superset superset/superset \
  -f examples-values.yaml

# One-shot: dashboard -> subagents
datus-agent bootstrap-bi --database superset

MCP protocol

MCP Protocol — client & server

Datus speaks Model Context Protocol in both directions. Consume any external MCP server, or expose Datus's own database and context-search tools to Claude, Cursor and other MCP hosts.

MCP Client

Wire external MCP tools into the agent from the CLI: .mcp add with stdio, http or sse transports.Config lives in ~/.datus/conf/.mcp.json.

MCP Server

Run datus-mcp in static or dynamic mode. Static serves one namespace, dynamic routes multiple namespaces by URL path.Exposes 8 database + 8 context-search tools out of the box.

Full MCP interface details in the docs

Skills

Agent Skills & Marketplace

Modular capability packs (v0.2.5) following the agentskills.io spec. Discover, install and publish skills from the built-in marketplace.

Bash Skills

Shell scripts guarded by an allow-list of commands; safe to expose to the agent.

Function Skills

Python callables loaded through load_skill() and invoked as tools.

Isolated Subagent Skills

Run in a forked subagent context (Explore / Plan / general-purpose) with its own scratchpad.

terminalbash

# Authenticate to the marketplace
datus skill login --marketplace http://datus-marketplace:9000

# Discover & install
datus skill search sql
datus skill install sql-optimization

# Publish your own
datus skill publish ./skills/my-skill --owner murphy

Observability

Observability & optional tools

Trace every LLM call, augment platform docs with web search, or debug prompts locally as YAML.

observability.envenv

Tool	Purpose	Configuration
LangSmith	LLM call tracing & debugging	`LANGSMITH_TRACING=true · LANGSMITH_API_KEY · LANGSMITH_PROJECT`
Langfuse	Agent + tool full-chain tracing (OTel / OpenInference)	`LANGFUSE_PUBLIC_KEY · LANGFUSE_SECRET_KEY · LANGFUSE_HOST`
LLM Trace	Local YAML dump of prompts and completions	`--save_llm_trace → {agent.home}/trajectory/`
Tavily	Web-search fallback for platform documentation	`TAVILY_API_KEY`
GitHub Token	Rate-limit-safe pull of platform docs from GitHub	`GITHUB_TOKEN`

FAQ

Frequently asked questions

Storage, embeddings, semantic layers, MCP, BI copilot, and how databases and models fit in.

Where do databases and models fit in?

Datus ships native database adapters covering the full range from SQLite to Snowflake, plus first-party LLM providers and any OpenAI-compatible endpoint. This page covers everything else you can plug in — storage backends, embeddings, semantic layer, BI copilot, MCP, skills, and observability.

Can I run Datus entirely on my laptop?

Yes. The default storage backend is LanceDB (vector) plus SQLite (relational), with zero configuration required. Data is written to data/datus_db_<namespace>/. Swap to PostgreSQL (pgvector) or Milvus as your deployment grows, without changing application code.

Is MetricFlow the only supported semantic layer?

MetricFlow is the only shipped adapter today, but the semantic layer is plugin-architected via Python entry points. You can register your own adapter alongside datus-semantic-metricflow through [project.entry-points."datus.semantic_adapters"] and expose it to the agent's context.

Which BI platforms are supported for Dashboard Copilot?

Apache Superset is supported today. Datus reads a dashboard, extracts every chart's SQL, builds a semantic model and emits two subagents — GenSQL for querying and GenReport for attribution and root-cause. Tableau, PowerBI and Looker adapters are on the roadmap.

What's the difference between MCP Client and MCP Server?

MCP Client lets Datus consume external MCP tools over stdio, http or sse transports. MCP Server exposes Datus's own database and context-search tools to any MCP-compatible host — Claude Desktop, Cursor, or another agent — with 8 database plus 8 context-search tools available out of the box.

Bring Your Stack, We Plug In

Databases, models, semantic layers, BI copilots and observability — every layer is a plugin.

Get started free Read the docs Browse adapters on GitHub