Vanilla Agents

$1,200 USD per install.

A set of independent libraries plus a scaffold project for building production-ready AI agents, fast. Roughly 60–70% of any agent codebase is infrastructure plumbing: structured logging, OpenTelemetry, LLM clients, FastAPI middleware, session management, MCP tool integration, streaming lifecycle. Vanilla Agents ships that plumbing as composable libraries, so your team writes business logic on day one, not week six.

It comes with an SDK-agnostic runtime interface. You start with the Claude Agent SDK, and you can swap in OpenAI Agents, Bedrock or another runtime without rewriting your service layer.

What you get

• 3 standalone infrastructure libraries: structured logging (with data sanitization and session aggregation), OpenTelemetry (settings, exporters, decorators) and LLM clients (an OpenAI-compatible factory)
• 1 agent toolkit library: settings, middleware, request/response models, session bridge, generic runtime interface, SDK-agnostic MCP tool system, skills system and a base agent service
• 1 scaffold project that assembles everything into a runnable agent: FastAPI app, health checks, an /invocations endpoint, example MCP tools, an example skill, Dockerfile, docker-compose and CI/CD workflows
• AWS Bedrock AgentCore deployment templates

Tools / stack

• Python + Pydantic + FastAPI: service layer
• structlog + OpenTelemetry: logging and observability
• httpx (OpenAI-compatible): LLM client
• Claude Agent SDK: default runtime (swappable through the AgentRuntime interface)
• AWS SSM Parameter Store: config in deployed environments
• AWS Bedrock AgentCore: deployment target

Why “vanilla”

You get the plumbing, not opinions about your business logic. Each agent extends 6 abstract methods (get_runtime, get_system_prompt, get_mcp_servers, get_allowed_tools, get_skills, _model_name) and the toolkit handles the rest: aggregator setup, streaming, interaction capture, SSE formatting, session persistence, error handling and cleanup. The same pattern, proven in production agents.