Atmosphere/atmosphere

What it solves

Atmosphere provides a real-time engine for deploying AI agents as production-grade services on the JVM. It solves the challenge of moving agents from simple demos to enterprise environments by providing built-in streaming transports, strict governance (like PII redaction and cost ceilings), durable session management for long-running tasks, and the ability to swap AI frameworks without rewriting endpoints.

How it works

Atmosphere acts as a service layer that wraps around AI runtimes. It uses a broadcaster pipeline to stream tokens from an LLM to clients via WebSocket, SSE, gRPC, or long-polling. Developers use the @Agent annotation to define agents, which can be "deep agents" with built-in memory, planning tools, and virtual filesystems. It employs a Runtime SPI (Service Provider Interface) with multiple adapters (e.g., Spring AI, LangChain4j, Anthropic) to remain portable across different AI frameworks.

Who it’s for

It is designed for teams building human-in-the-loop, multi-channel, governed AI agents within an existing JVM stack (such as Spring Boot, Quarkus, or Netty) who require production-level stability and oversight.

Highlights

• Multi-Transport Streaming: Native support for WebSocket, SSE, gRPC, and long-polling out of the box.
• Enterprise Governance: Includes policy admission, @AgentScope for prompt blocking, PII rewriting, and admin kill switches.
• Durable HITL: Human-in-the-loop approvals that hibernate without holding threads and resume via REST endpoints.
• Extensive Integration: 12 runtime adapters for various frameworks and native clients for Anthropic and Cohere.
• Deep Agent Capabilities: Built-in support for memory, planning (write_todos), and a virtual filesystem.
• Broad Protocol Support: Exposes agents via MCP, A2A, AG-UI, and various chat channels like Slack, Telegram, and Discord.