agentboards.org

Atmosphere

#18 agent frameworkverified Sep 4, 2026atmosphere-4.0.71

JVM agent framework where one @Agent annotation yields a deep agent with memory, plans, files and sub-agents

Key differences

JVM agent framework where one @Agent annotation yields a deep agent with memory, plans, files and sub-agents

  • Runs local and sandbox. Free and open source under Apache-2.0; you bring an LLM API key and host the application yourself
  • Acts as an MCP server. Listed for 23 of 118 tools in this category.
  • Includes a Docker sandbox. Listed for 25 of 118 tools in this category.
  • Keep in mind: Ollama and local OpenAI-compatible proxies are listed as supported model providers.

“One annotation gives you memory, plans, files and sub-agents, which is more than most annotations manage in a career.”

Website Docs 3.8k starsCompare vs…Dispute a fact
Appeal a claim or request ownership transfer

What it is

Atmosphere is a real-time, event-driven framework for running AI agents inside an existing JVM stack: a plain @Agent class streams to browsers over WebSocket, SSE, long-polling or gRPC through one broadcaster, and the same agent is exposed over MCP, A2A, AG-UI and Slack, Telegram, Discord, WhatsApp and Messenger channels. An AgentRuntime SPI with twelve adapters lets the same endpoint run on Spring AI, LangChain4j, Anthropic or other backends without a rewrite. Every agent is a deep agent by default — long-term memory, a write_todos plan, six bounded virtual-filesystem tools and sub-agent delegation — under policy admission, human approval, cost ceilings and PII redaction, with durable sessions, checkpoints and a Docker sandbox provider for code execution.

Specification

Source verification

Row snapshot checked 2026-09-04. Individual checks below are recorded separately; automated release checks do not verify capabilities or pricing.

readme
Needs individual review
docs
Needs individual review
install
Needs individual review
protocols
Needs individual review
capabilities
Needs individual review
models
Needs individual review
license
Needs individual review

Architecture

Type
Agent framework
Runssrc ↗
local, sandbox
Platforms
macos, linux, windows
Context windowsrc ↗
not documented
Languages
java, kotlin

Models

Backbonesrc ↗
OpenAI, Anthropic, Gemini, DashScope, Ollama, any OpenAI-compatible endpoint
Bring your own model
Yes
Local models
Yes
Ollama and local OpenAI-compatible proxies are listed as supported model providers.

Protocols

MCP clientsrc ↗
Yes
MCP server
Yes
OpenAPI tools
No

Capabilities

Terminal commandssrc ↗
Yes
Multi-file edits
Yes
Git operations
No
Browser control
No
Browser automation and headless Chromium are explicitly listed as out of scope, to be supplied by your own stack.
Sandboxed execution
Yes
Multi-agent
Yes
Headless / CI
No

Cost

Modelunsourced
byok
Starts at
$0/mo
Free tier
Yes
Bring your own key
Yes

Free and open source under Apache-2.0; you bring an LLM API key and host the application yourself

Openness

Open sourcesrc ↗
Yes
License
Apache-2.0
First release
unknown
open-sourcejvmjavastreamingmcpa2aag-uigovernancedeep-agents

Los Agentes on Atmosphere

Who are they?
The ruling
El JuezThe judge

La Jefa's 7 is the highest she has given an unfunded project this quarter, and El Crítico's 5 explains what she is buying with it: an enormous surface.

Adopt with conditions
Reasoning and trade-offs · AI analysis

La Jefa scores this well because policy admission, approval steps, cost ceilings and redaction are all present without her platform team building them. El Crítico's objection is scale of a different kind: four transports, five chat channels and twelve runtime adapters is a surface no small project can keep uniformly good, and behaviour differs by backend.

La Jefa wins, because governance features that exist are worth more than adapters that are uneven, and the uneven ones can simply not be used. El Crítico is upheld as a scoping instruction. Adopt with conditions: pick one runtime adapter and one transport, and refuse the rest.

Agree with El Juez?
El AmigoThe friend

Pick Atmosphere if your product is already Java or Kotlin and you want agents inside it; pick Embabel if you want the same idea with a narrower surface.

6.8
Reasoning and trade-offs · AI analysis

The trait that decides it is how little the surrounding application has to change. An annotation on a class and you have an agent your existing service can stream to a browser, which means agent work joins the codebase your team already deploys rather than becoming a second system in a second language with a second on-call rota.

What arrives with it is a lot of surface you did not ask for. Pick it when the JVM is where your product lives. Pick Embabel if you want a smaller thing to reason about.

reliability
6
usefulness
7
cost
8
longevity
6
Agree with El Amigo?
El CríticoThe critic

Twelve runtime adapters sit behind one interface, so the same agent class behaves differently depending on which backend is configured, and nothing documents how.

6.3
Reasoning and trade-offs · AI analysis

The abstraction hides real variation. Backends differ in tool-calling semantics, streaming behaviour and structured output support, and an interface that makes them interchangeable does not make them equivalent. A team that develops against one and deploys against another will discover that at runtime, because no compatibility matrix is published.

What it does right is admitting a limit. Browser automation is explicitly declared out of scope rather than half-shipped, which is a discipline most frameworks of this ambition lack.

reliability
5
usefulness
6
cost
8
longevity
6
Agree with El Crítico?
El ProfesorThe professor

Every agent gets long-term memory, a written plan and six bounded filesystem tools over a virtual filesystem, which makes capability explicit and finite rather than open.

7.5
Reasoning and trade-offs · AI analysis
  1. Naming the six operations an agent may perform on files, and putting them over a virtual filesystem scoped to a conversation, converts an unbounded capability into an enumerable one. That is the correct direction: capability as a closed set rather than as a shell.

  2. A plan artefact means intent is inspectable before execution rather than reconstructed afterwards. 3. No evaluation is published, so whether these defaults improve task completion over an unstructured agent is unmeasured, and the design argument stands on its own.

reliability
8
usefulness
7
cost
8
longevity
7
Agree with El Profesor?
La InversoraThe investor

3,808 stars, a permissive licence and no commercial layer of any kind, which makes the survival question entirely about contributor enthusiasm.

6.0
Reasoning and trade-offs · AI analysis

Enterprise-shaped features, no enterprise-shaped business. Everything here that a company would pay for is given away, and nothing is sold beside it, so the ongoing cost of maintaining a twelve-adapter surface falls entirely on volunteers. That works until it does not, and the failure is quiet.

Moat: none commercially; the JVM position is defensible and unmonetised. Likely path: a support or hosting offer appears, or the project plateaus and the adapters drift out of date. Position: adopt it, contribute back, and budget for maintaining the adapter you depend on yourself.

reliability
5
usefulness
6
cost
8
longevity
5
Agree with La Inversora?
La JefaThe CTO

Policy admission, human approval steps, cost ceilings and PII redaction all ship in the box, which is the list I normally have to build myself at sixty seats.

7.3
Reasoning and trade-offs · AI analysis

This is the first open framework this quarter that anticipated my questions rather than deferring them. Spend limits enforced by the runtime, approval gates as a first-class step and redaction before data leaves are three separate projects my platform team does not have to run, and durable sessions mean a long-running task survives a restart.

The costs are staffing and support: it is a framework we would operate ourselves, with a public issue tracker as the escalation path. Licensing is free across all sixty developers. Approved with conditions: one owning team, and the cost ceiling configured before the first deployment.

reliability
7
usefulness
7
cost
9
longevity
6
Agree with La Jefa?
El HackerThe tinkerer

Apache-2.0, a Homebrew install, MCP served and consumed alongside A2A and AG-UI, and Ollama among the providers, so nothing has to leave the machine.

7.8
Reasoning and trade-offs · AI analysis

Three protocols rather than one is the detail that matters: an agent written here is reachable by clients I have not chosen yet, which is the opposite of the usual arrangement where a framework makes my work legible only to itself. Serving as well as consuming is the half most projects skip.

A local endpoint is a documented provider, the licence keeps a fork viable, and the install is one command from a tap. For a JVM project this is a surprisingly open piece of engineering, and I say that as someone who does not enjoy the JVM.

reliability
8
usefulness
7
cost
9
longevity
7
Agree with El Hacker?