DEV.co
MCP Servers · Atmosphere

atmosphere

Atmosphere is a Java framework that lets you build AI agents once and deploy them across WebSocket, SSE, gRPC, and HTTP/3, with built-in memory, planning, and governance. It abstracts transport and runtime adapters so the same agent code works with OpenAI, Anthropic, LangChain4j, Spring AI, and a dozen other backends.

Source: GitHub — github.com/Atmosphere/atmosphere
3.8k
GitHub stars
759
Forks
Java
Primary language
Apache-2.0
License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
RepositoryAtmosphere/atmosphere
OwnerAtmosphere
Primary languageJava
LicenseApache-2.0 — OSI-approved
Stars3.8k
Forks759
Open issues7
Latest releaseatmosphere-4.0.61 (2026-07-07)
Last updated2026-07-08
Sourcehttps://github.com/Atmosphere/atmosphere

What atmosphere is

JVM-based real-time transport and orchestration layer for stateful AI agents. Provides `@Agent` annotation, `AgentRuntime` SPI with 12 adapters, event-sourced coordination, durable session persistence (SQLite/Redis), policy admission, HITL workflows, and multi-channel protocol support (MCP, A2A, AG-UI, Slack/Discord/Telegram).

Quickstart

Get the atmosphere source

Clone the repository and explore it locally.

terminalbash
git clone https://github.com/Atmosphere/atmosphere.gitcd atmosphere# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Multi-channel AI agents inside existing JVM apps

Stream agent state over WebSocket/SSE/gRPC to web/mobile clients, Slack bots, and MCP-compatible tools from a single Spring Boot or Quarkus codebase without rewriting endpoints per channel.

Human-in-the-loop workflow automation

Build governed, pausable agent runs with durable HITL approvals that hibernate without holding threads, persist state, and resume across container restarts using `Workflow<S>` and `CheckpointStore`.

Model-agnostic agent applications

Swap LLM providers (OpenAI → Anthropic → Ollama) and runtime adapters (Spring AI ↔ LangChain4j) via classpath modules without touching agent code; contract-tested capability flags prevent runtime surprises.

Implementation considerations

  • Transport layer runs on your JVM host (Tomcat, Jetty, Netty, Undertow, Spring Boot, Quarkus, any servlet container); no separate Atmosphere-managed infrastructure, but scheduling and compute are your responsibility.
  • Persistence for durable sessions and HITL workflows requires SQLite or Redis (`atmosphere-durable-sessions{-sqlite,-redis}` modules); design CheckpointStore and ConversationPersistence implementations or use defaults.
  • Governance (policy admission, cost ceilings, PII redaction) integrates inline on the critical path; test authorization and approval workflows before production deployment.
  • Model providers are pluggable via `AgentRuntime` SPI; vet capability flags of your chosen adapter (tool calling, streaming, structured output, prompt caching) early—not all adapters support all features.
  • Multi-channel endpoints (MCP, A2A, Slack, Discord, etc.) are opt-in modules; include only what you need to keep classpath lean and security surface minimal.

When to avoid it — and what to weigh

  • Serverless, stateless agent functions — Atmosphere optimizes for long-lived stateful sessions. If agents should spin down when idle or run as bursty request-response functions, a dedicated agent platform (AWS Bedrock Agents, Vertex AI Agents) is a better fit.
  • Non-JVM tech stack — Atmosphere is Java-native. Python/Node.js/Go teams will find adoption friction; the JavaScript SDK is client-only, not a backend runtime.
  • Minimal operational overhead required — Durable sessions, governance, and HITL features require persistent stores (SQLite, Redis) and careful authorization topology. Teams seeking zero-config inference should look elsewhere.
  • Real-time browser automation at scale — The sandbox SPI supports Docker-based code execution, but no built-in browser pool or Chromium scaling; headless automation requires external integration.

License & commercial use

Apache License 2.0 (Apache-2.0). Permissive OSI license: allows commercial use, modification, and distribution with attribution and liability disclaimer. No copyleft or patent retaliation clauses.

Apache-2.0 is explicitly permissive for commercial use. No license restrictions on building proprietary products with Atmosphere. Consult counsel if incorporating modified Atmosphere code; license obligations flow from Apache-2.0 terms (attribute, include license in distributions). Unknown: presence of any transitive GPL or restrictive dependencies; review `pom.xml` and dependency tree in your build.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Atmosphere provides governance primitives (policy admission, `@AgentScope`, plan-and-verify, cost ceilings, PII rewriting, durable HITL approvals, admin kill switches) but does not itself guarantee security. Critical considerations: (1) Authorization topology must be designed and tested; policy admission runs on critical path but is only effective if policies are sound. (2) HITL and approval workflows are durable but require secure REST surface and audit logging. (3) Code execution via `SandboxProvider` (Docker default) requires sandboxed container runtime and network isolation. (4) Transitive dependency risk: Apache-2.0 is permissive, but build must be scanned for GPL/restrictive licenses in dependency tree. (5) Session persistence (SQLite/Redis) must be encrypted and access-controlled. (6) Model provider secrets (LLM_API_KEY) must be injected securely, not hardcoded. (7) No built-in rate limiting or DDoS mitigation; rely on transport layer (servlet container, gRPC server) and reverse proxy.

Alternatives to consider

LangChain Python / LangChain.js

Multi-language agent framework with broader ecosystem integration (LangChain community tools, LangSmith observability). Stateless by default, serverless-friendly. Better for non-JVM stacks or minimal governance needs. Lacks Atmosphere's real-time transport abstraction and durable HITL workflows.

Anthropic Agents (AWS Bedrock, Claude API)

Fully managed, bursty agent platform with built-in tool use, long-context, and Claude models. Zero-ops infrastructure; pay-per-invocation. Better for stateless autonomous agents. No governance, HITL, or multi-channel streaming; lock-in to Anthropic models.

Vertex AI Agents (Google Cloud)

Managed agent orchestration with built-in web search, code execution, and multi-turn grounding. Integrated with GCP services. Stateless, event-driven. Better for Google Cloud-native stacks. No transport abstraction, no durable HITL, no model flexibility.

Software development agency

Build on atmosphere with DEV.co software developers

Atmosphere bridges the gap between LLM APIs and production agent infrastructure. If you're building long-lived, human-approved agents inside a JVM stack, explore the quick-start guide and CLI tooling to see if Atmosphere fits your needs.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

atmosphere FAQ

Do I need to run Atmosphere as a separate service?
No. Atmosphere is a library you embed in your JVM app (Spring Boot, Quarkus, etc.). It runs in-process and delegates compute/scheduling to your host. No separate Atmosphere infrastructure.
Can I use Atmosphere with my favorite LLM provider?
Yes, through one of 12 runtime adapters (Built-in, Spring AI, LangChain4j, Google ADK, Embabel, Koog, Semantic Kernel, AgentScope, Spring AI Alibaba, Anthropic, Cohere, CrewAI). Each adapter implements `AgentRuntime` with contract-tested capability flags. Check your provider's adapter before deploying.
How do I pause an agent for human approval?
Use durable `Workflow<S>` over `CheckpointStore` and `@AgentScope` with HITL approval gates. Workflows hibernate without holding threads, persist state, and resume across container restarts. Approvals are exposed via REST approval surfaces.
Is Atmosphere suitable for stateless, bursty inference?
No. Atmosphere optimizes for long-lived stateful agent sessions. For stateless request-response agents that scale to zero, use AWS Bedrock Agents, Vertex AI Agents, or Anthropic's API directly.

Work with a software development agency

Adopting atmosphere is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate mcp servers software in production.

Deploy governed AI agents across web, mobile, and enterprise channels.

Atmosphere bridges the gap between LLM APIs and production agent infrastructure. If you're building long-lived, human-approved agents inside a JVM stack, explore the quick-start guide and CLI tooling to see if Atmosphere fits your needs.