A framework for building streaming AI agents on the JVM with a built-in transport layer, token-level governance gateway, multi-protocol distribution, multi-channel messaging, and multi-agent orchestration.
Positioning#
Atmosphere is a JVM-based streaming AI agent framework that decouples LLM inference via AgentRuntime SPI, adapting to 9 external AI runtimes (OpenAI-compatible, Spring AI, LangChain4j, Google ADK, JetBrains Koog, Microsoft Semantic Kernel, Alibaba AgentScope, Embabel, Spring AI Alibaba). Runs on Spring Boot 3.5/4.0, primarily Java (Maven), with a JavaScript/TypeScript client library atmosphere.js.
Transport Layer#
Supports five transport protocols: WebTransport/HTTP3 (QUIC), WebSocket, SSE, Long-polling, gRPC. Auto-negotiation via AsyncSupport, with Jetty 12 QUIC native or Reactor Netty HTTP/3 sidecar. Reconnection via X-Atmosphere-Run-Id header with mid-stream buffer replay.
Agent Development Model#
Annotation-driven: @Agent auto-registers endpoints/tools; @Prompt for streaming messages; @Command for slash commands (with confirm human-in-the-loop); @AiTool registers tools and auto-exposes as MCP tools. Protocol switching via Maven dependencies without code changes.
Protocol Interoperability#
Three Agent protocols: MCP (Model Context Protocol), A2A v1.0.0 (with Agent Card discovery), AG-UI. Modules auto-register endpoints: /atmosphere/agent/{name}/mcp, /atmosphere/agent/{name}/a2a, /atmosphere/agent/{name}/agui.
Multi-Channel Distribution#
A single @Command dispatches to Slack, Telegram, Discord, WhatsApp, Messenger via atmosphere-channels module.
Multi-Agent Orchestration#
@Coordinator, @Fleet, @AgentRef annotations supporting local (in-JVM) and A2A (HTTP JSON-RPC) transport with parallel/sequential/conditional routing and coordination logs.
Memory & Sessions#
Sliding window and LLM summarization memory with SQLite/Redis persistence.
Governance & Compliance#
- Guardrails: PII redaction, output length anomaly detection, cost ceilings
- Policy Engine: GovernancePolicy SPI (YAML, compatible with Microsoft Agent Governance Toolkit schema), built-in AllowList/DenyList/MessageLength/RateLimit/ConcurrencyLimit/TimeWindow/Authorization/ConfidenceThreshold
- Ops Controls: KillSwitch, DryRun, SwappablePolicy hot-reload, SloTracker
- Compliance Evidence: OWASP Agentic Top 10 + EU AI Act/HIPAA/SOC 2 matrix, CI-enforced
- Advanced Verification: Plan-and-verify (static verification of LLM tool-call workflows), Commitment Records (W3C Verifiable-Credential subtype, Ed25519 signatures)
Sandbox#
SandboxProvider SPI with built-in DockerSandboxProvider (default --network none) and InProcessSandboxProvider, open ServiceLoader for Firecracker/Kata/Vercel Sandbox/E2B/Modal/Blaxel.
Admin Console#
/atmosphere/admin/ (Vue UI) + /api/admin/* (REST) with Triple-gate authorization, Flow Viewer topology, and governance dashboard.
Quick Start#
brew install Atmosphere/tap/atmosphere
atmosphere run spring-boot-multi-agent-startup-team
atmosphere new my-agent --template ai-chat