Solon AI: a modular Java agent, RAG and MCP framework that runs from Java 8 and inside Spring Boot, Vert.x or Quarkus
Java AI application development framework (supports LLM-tool,skill; RAG; MCP; Agent-ReAct,Team-Agent). Compatible with java8 ~ java26. It can also be embedded in SpringBoot, jFinal, Vert.x, Quarkus, and other frameworks.
At a glance
- What is it?
- opensolon/solon-ai is the AI subproject of the Solon Java framework, offering a unified ChatModel, RAG components, MCP server and client support, and ReAct and team agents. It splits into more than twenty modules, tracks a pinned MCP revision, and ships frequent releases.
- Who is it for?
- Pick Solon AI if you build AI features on the JVM, including older Java versions, and want ChatModel dialects, RAG, MCP and agents from one family of modules. Look at other options if you need a Java framework outside the Solon ecosystem conventions, or if you want a single dependency with no module choices.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 2 days ago.
- What is it written in?
- Mainly Java, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- opensolon/solon-ai473 ★ · +18
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| opensolon/solon-ai | 455 | 473 |
18 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
It positions itself beside LangChain, LangGraph and LlamaIndex, for the JVM
The README opens with a direct comparison: Solon AI is the same type of development framework as LangChain, LangGraph and LlamaIndex, for Java. Its tagline lists LLM with tools and skills, RAG, MCP and Agent with ReAct and Team modes, and its motto is restraint, efficiency and openness.
The description makes a compatibility claim: it works with Java 8 through Java 26, and it can be embedded in SpringBoot, jFinal, Vert.x, Quarkus and other frameworks. The README also says it fits into the Solon ecosystem, and it lists embedding examples for third party frameworks in a separate solon-ai-mcp-embedded-examples repository.
The consequence is that you are not forced onto Solon's own web layer. A team on Spring Boot can add AI components without a platform migration, though the framework's idioms, such as annotation names for MCP endpoints, come from Solon.
More than twenty directories mean you choose modules, not one jar
The top level shows how the project is cut up. Core pieces include solon-ai-core, solon-ai-dialects, solon-ai-agent, solon-ai-flow, solon-ai-mcp, solon-ai-talents, solon-ai-router, solon-ai-harness, solon-ai-loop, solon-ai-sandbox and solon-ai-ui. Protocol work sits in solon-ai-acp, solon-ai-anp, acp-core, acp-annotations, acp-agent-support, mcp-core and mcp-json-jackson2. RAG is split three ways: solon-ai-rag-loaders, solon-ai-rag-repositorys and solon-ai-rag-searchs. There are also solon-ai-parent and a pom.xml at the root.
The practical result is a dependency decision. You add the modules your feature needs, for example core plus a dialect and the MCP module, rather than pulling everything. The cost is that you must learn which module owns which capability, and the overview text does not map features to artifact names, so check the module directory before you add a dependency.
ChatModel hides vendor differences, but you must name the dialect
ChatModel is the general purpose LLM call interface. It supports synchronous and reactive calls and has built-in dialect adaptation, Tool, Skill and ChatSession. The dialects listed are OpenAI, Gemini, Claude, Ollama, DeepSeek and Dashscope, so the same code can talk to different vendors.
The example builds a model against an Ollama URL at http://127.0.0.1:11434/api/chat, with the provider set to ollama and the model qwen2.5:1.5b. The comment beside provider is the useful part: you need to specify the vendor, because it identifies the interface style, also called the dialect.
That is where a mistake hurts. Point the builder at an endpoint with the wrong provider name and the request shape will not match what the server expects. The unified interface removes vendor specific client code, not the need to know which API style your endpoint speaks. Tools are attached per call through options, as in the weather tools example.
The same builder also takes a default talent, shown with McpGatewayTalent, so a model can be given a standing capability at construction time instead of per request. That is convenient for a shared model instance, but it means every call on that instance carries the talent, so build separate models when different requests need different capabilities.
Talents activate by prompt content and change their instructions per user
Talents are the project's name for skill-like capabilities. The example defines a talent called order_expert with the description Order Assistant. Two hooks stand out. isSupported takes the prompt and decides admission: in the example the talent is activated only when the user message contains the word order. The instruction hook builds text dynamically: if the prompt metadata user_level equals VIP, it injects an instruction to call fast_track_tool first.
This makes behaviour depend on the content of each request. A talent is not always on, so a message that does not trigger the admission test will never see its instructions or tools.
The consequence is a testing burden. Admission by substring, as in the example, is easy to write and easy to miss: a customer who writes about a purchase without the word order bypasses the talent. If you rely on talents for business rules, test the admission conditions as carefully as the instructions themselves.
The RAG chain covers loading, splitting, embedding and reranking
RAG support is described as full link: DocumentLoader, DocumentSplitter, EmbeddingModel and RerankingModel. The example builds an EmbeddingModel with an API URL, key, provider and model plus batchSize(10), builds a RerankingModel the same way, then creates an InMemoryRepository on top of the embedding model.
Documents enter through a loader, in the example a PdfLoader given a URI, and are inserted with repository.insert. Retrieval is repository.search(query) returning a list of Document. If you want better ordering, rerankingModel.rerank(query, docs) rearranges them before the prompt is enhanced.
Two cautions follow. The example repository is InMemoryRepository, so as written the knowledge base lives in process memory and is rebuilt on restart. Persistent stores belong to the solon-ai-rag-repositorys modules, which you must choose and configure separately. Reranking is an extra model call per query, so it adds latency and cost, which is why the example presents it as optional.
MCP support targets one protocol revision and uses annotations for servers
The README claims deep integration with the MCP protocol at revision MCP_2025_06_18, with cross platform sharing of tools, resources and prompts. A server is an annotated class, and the client is a builder:
//server
@McpServerEndpoint(channel = McpChannel.STREAMABLE, mcpEndpoint = "/mcp")
public class MyMcpServer {
@ToolMapping(description = "Checking the weather")
public String getWeather(@Param(description = "city") String location) {
return "It's sunny, 25 degrees";
}
}On the client side, McpClientProvider.builder() takes a channel of McpChannel.STREAMABLE and a url such as http://localhost:8080/mcp.
The pinned revision is the thing to check. A client or server built against a different protocol date may not interoperate with this one, and the example uses the streamable channel, so a counterpart that only speaks another transport needs a different channel setting. The embedded examples repository shows hosting MCP inside other frameworks.
Agents turn reasoning into graph flows, with ReAct and team modes
The Agent module is described as transforming reasoning logic into graph driven collaboration flows. The README states this lets agent reasoning become an observable and governable computation flow graph, which is the selling point against free form loops.
A ReActAgent is created from a chat model with a name, description and default tools, for example a weather expert given a weather tool, and the comment on the call describes the cycle: Think, Call tool, Observe, Summarize. SimpleAgent is named as an alternative. A team agent assigns member roles automatically, according to the example comment.
Two larger projects are offered as production ready references: SolonCode, a Java implementation of Claude Code, and SolonClaw, a Java implementation of OpenClaw. They show the framework on real agent products.
The cost of graph orchestration is that behaviour becomes more structured and therefore more explicit to debug, but also more to learn than a plain loop.
Releases arrive roughly monthly, and the changelog is the thing to read
The three newest tags are v4.0.6 on 2026-08-17, v4.1.0 on 2026-09-07 and v4.1.1 on 2026-10-01, and the last push was on 2026-11-02. The repository is not archived and is licensed Apache-2.0. The homepage is solon.noear.org, with a learning article at solon.noear.org/article/learn-solon-ai.
A minor version bump from 4.0 to 4.1 inside three weeks suggests the API is still moving, so a pin to a specific tag matters more than a loose range. UPDATE_LOG.md and TEST.md sit at the top level, and the changelog is where behaviour changes will be recorded.
The project also hosts mirrors on Gitee and GitCode, so the same code appears in three places. Check that the repository you clone matches the version in your build file. The README is available in English and in Chinese.
The use cases the project lists are broad: autonomous agents, RAG assistants, multi agent orchestration, controlled business workflows such as approval flows, document processing, text to SQL dashboards, automated testing and low code AI workflow platforms. Those are categories it says you can build, each named with an example product from elsewhere, not features that ship ready made in this repository, so treat the list as scope, not as a feature inventory.
Editorial conclusion
Pick Solon AI if you build AI features on the JVM, including older Java versions, and want ChatModel dialects, RAG, MCP and agents from one family of modules. Look at other options if you need a Java framework outside the Solon ecosystem conventions, or if you want a single dependency with no module choices. Before adopting, check the MCP protocol revision it targets and read UPDATE_LOG.md for breaking changes between the 4.0 and 4.1 lines.
Frequently asked questions
What is Solon AI used for?
Solon AI is a Java framework for LLM calls, RAG knowledge bases, MCP servers and clients, and ReAct or team agents, in the same family as LangChain and LlamaIndex.
Which Java versions and frameworks does Solon AI support?
The description says it is compatible with Java 8 through Java 26 and can be embedded in SpringBoot, jFinal, Vert.x, Quarkus and other frameworks.
Which LLM providers does Solon AI's ChatModel support?
ChatModel adapts dialects for OpenAI, Gemini, Claude, Ollama, DeepSeek and Dashscope, and you name the provider when building the model.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/opensolon-solon-ai)Star history chart for your README
[](https://hysenlabs.com/projects/opensolon-solon-ai)Was this useful?