Pillar — SDK

The Python library that makes adding AI to your product feel like a normal pull request.

The SDK pillar is the pip-installable surface — pip install sagewai and you have UniversalAgent, tool calling, memory, workflows, directives, MCP, sandbox, and model-swap. No admin server required. No other Sagewai pillar required. The platform layers on top when you need them; the SDK is useful on its own.

What the pillar does

• UniversalAgent — the core agent class. Single-turn or multi-turn, tool-calling or pure chat, any LLM via LiteLLM.
• Tools and MCP — Python tools registered with the agent, and MCP-compliant tool servers (Anthropic's open standard) attached transparently.
• Memory and RAG — vector memory, graph memory (GraphMemory, NebulaGraphMemory), strategy-based extraction, semantic-checkpoint recall.
• Workflows — DurableWorkflow for multi-step agent compositions with persistence.
• Directives — @datetime, @context, @memory, /tool.name directives that expand at prompt-resolution time. The harness any LLM moat.
• Sandbox — agents running inside isolated containers with scoped credentials. The Sealed-spine integration for the SDK.
• Model swap — same code, any LLM. Anthropic, OpenAI, Ollama, LM Studio, Modal, Vast.ai endpoints — all behind the same UniversalAgent constructor.

What proves it works

Primary lighthouse

Moderation and classification — the cleanest single proof of ML and LLM both first-class. Three HuggingFace classifiers in a sealed sandbox-ml container, judged by a Haiku-class LLM, audit per call.

Sibling lighthouses

• Memory and retrieval — semantic checkpoint recall (Example 37) and graph-beats-vector for incident response (Example 41).

Foundation examples

The SDK pillar's foundation surface is most of the Foundation section. Pick the ones that match what you're shipping:

• 02 — tool_agent — tool calling
• 04 — memory_agent — memory basics
• 05 — workflow — multi-step workflows
• 06 — guardrails — safety filters
• 07 — mcp_tools — MCP integration
• 08 — directives — directive library
• 03 — multi_model, 13 — model_routing, 18 — local_llm_routing — model swap

Pattern examples

• 19 — domain_model
• 21 — full_stack
• 27 — app_factory
• 29 — memory_strategies
• 31 — grounded_multi_model
• 32 — global_shared_memory

Where to go to ship it

• Quickstart — the 60-second hello-world.
• Getting started — the longer "build a portfolio site with Claude Code" tour.
• Core concepts: Agents — UniversalAgent, composition patterns.
• Core concepts: Memory & RAG — the memory subsystem.
• Core concepts: Workflows — DurableWorkflow, approval gates.
• Core concepts: Directives — directive-library deep dive.
• API reference: Agents — the full agent API.

Related

• Sealed spine — the security architecture that makes the SDK safe to point at production.
• Pillars overview — the other four pillars and how they compose with the SDK.