How mastra-ai/mastra works
Mastra is a TypeScript framework for building and operating AI applications. Developers define agents for open-ended tasks or workflows for explicit multi-step control; agents can use models and tools, with context and memory supporting execution. Human approval and persisted state allow suspended work to resume. Applications expose capabilities through server and MCP integrations, while SDKs and UI packages connect clients. The README documents evaluation and observability for production refinement. Runtime wiring beyond the sampled excerpts is represented only where the README establishes the capability; the excerpts do not establish detailed call paths.
Main entry points: core/src Agents · server/src Server endpoints · session-run-engine.ts Human approval · search-engine.ts Workspace search
Components 20
AI Runtime
- Agents
packages/core/src - Workflows
packages/core/src - Model routing
packages/core/src - Tools
packages/core/src - Human approval
packages/core/src/agent-controller/session-run-engine.ts
Context and State
- Context and memory
packages/core/src - Workspace search
packages/core/src/workspace/search/search-engine.ts - Execution storage
packages/core/src
App Interfaces
- MCP servers
packages/mcp/src - Server endpoints
packages/server/src - Client SDKs
client-sdks/client-js/src - React integration
client-sdks/react/src - AI SDK integration
client-sdks/ai-sdk/src - Deployment
deployers
Production Operations
- Evaluations
packages/core/src - Observability
packages/core/src/observability/context-storage.ts
Other components
- Application developer
- End user
- Model providers
- Data sources
Connections 28
- Application developer → Agents: defines
- Application developer → Workflows: defines
- Agents → Model routing: requests inference
- Agents → Tools: invokes tools
- Workflows → Agents: orchestrates
- Workflows → Tools: runs steps
- Model routing → Model providers: routes requests
- Agents → Context and memory: uses context
- Context and memory → Workspace search: retrieves content
- Context and memory → Execution storage: persists state
- Workflows → Execution storage: persists execution
- Agents → Human approval: suspends for approval
- Workflows → Human approval: suspends for approval
- Human approval → Execution storage: resumes state
- Application developer → MCP servers: exposes resources
- MCP servers → Agents: serves agent capabilities
- Server endpoints → Agents: exposes runtime
- Server endpoints → Workflows: exposes runtime
- End user → Server endpoints: sends requests
- Client SDKs → Server endpoints: calls endpoints
- React integration → Client SDKs: uses SDK
- AI SDK integration → Server endpoints: connects UI
- Server endpoints → Deployment: runs on
- Agents → Evaluations: evaluated by
- Workflows → Evaluations: evaluated by
- Agents → Observability: traced by
- Workflows → Observability: traced by
- Workspace search → Data sources: indexes content