> ## Documentation Index
> Fetch the complete documentation index at: https://docs.praxa.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Add Praxa to an agent framework

> Use Praxa with Vercel AI SDK, OpenAI Agents, LangGraph, or another tool host without replacing your model or memory system.

Praxa is composable with an existing agent stack. Add only the contract you
need: governed tools through MCP, executable helpers through the SDK, or
read-only recall through memory federation.

```mermaid theme={null}
flowchart TB
  Host["Your agent framework"] --> Tools["Praxa tools"]
  Host --> Memory["Existing memory"]
  Tools --> Gateway["Integration Gateway"]
  Memory --> Adapter["Praxa read-only adapter"]
  Adapter --> Context["Provenance-rich recall"]
  Gateway --> Evidence["Run, event, receipt"]
```

## Framework paths

| Framework            | Best starting point                                | Preserved authority                                 |
| -------------------- | -------------------------------------------------- | --------------------------------------------------- |
| Vercel AI SDK        | [ToolLoopAgent tutorial](/tutorials/vercel-ai-sdk) | Host model loop plus Praxa server policy            |
| OpenAI Agents        | [OpenAI Agents tutorial](/tutorials/openai-agents) | Remote MCP and session memory remain separate       |
| LangGraph            | [LangGraph tutorial](/tutorials/langgraph)         | Long-term store adapted; checkpoints excluded       |
| Any MCP host         | [MCP host integration](/mcp/host-integration)      | Stable wire names, schemas, annotations, and scopes |
| Custom memory system | [Memory SDK guide](/sdk/memory)                    | Caller-owned clients and namespace resolution       |

## Migration strategy

Start with federation. Keep provider-owned memory in place, add a read-only
adapter, test two-tenant and two-subject isolation, and compare results. Move
data only if a later product requirement justifies migration and deletion is
covered end to end.

<Warning>
  A model can request a tool, but it cannot grant itself scope, choose the
  tenant, approve a consequential action, or declare an effect verified.
</Warning>