MySQL does not have a built-in "AI interface", so the first problem enterprises encounter when connecting to large models is figuring out where their data is. Shannon Boolean breaks this deadlock with the "in-situ upgrade" solution and officially launches its angel round financing.
The legacy MySQL lacks "AI interface", which has become the first threshold for enterprises to transform their data assets with AI capabilities
When enterprises connect their private data to large models, the first problem they encounter is often not the model itself, but the location of the data. In online transaction scenarios, MySQL has long been the default underlying infrastructure. According to public statistics, there are tens of millions of deployed instances worldwide, and hundreds of thousands of enterprises have incorporated it into their core technology stacks. However, the capability boundary of MySQL itself stays at transaction processing: it has no columnar execution for analytics, no model training and inference that can be directly called, no complete link of vector retrieval and Retrieval-Augmented Generation (RAG), nor any in-database safe-to-run Agent.
Neither of the two transformation paths formed thereby is easy. The first path is replacement, which migrates the online database to a new type of database with AI capabilities. In this scenario, data migration, SQL compatibility, operation and maintenance system as well as compliance audit need to be completely rebuilt, which sets an excessively high threshold for core systems that have been running for many years. The other path is external integration, which stacks vector database, big data platform, feature platform and large model gateway outside MySQL. This path brings small architecture changes, but data needs to be transferred back and forth between multiple systems, leading to link latency, duplicate storage costs, inconsistent permission standards, and security risks of context flowing outside the database. For industries with strict "data not leaving the database" requirements such as finance and government & public sectors, the compliance cost of the latter path is particularly prominent.
ShannonBase defines this blank area as the demand for "in-situ upgrade": without modifying business code or migrating data, the already running MySQL instances can acquire analytics, machine learning and Agent capabilities in place. The project is initiated by Shannon Data AI, a database kernel team, which adopts an open-source route, with code hosted on GitHub and available for trial now.
No replacement, only upgrade: columnar storage, inference and Agent runtime are all built into the kernel
ShannonBase is a set of open-source intelligent databases fully compatible with MySQL 8.4 LTS. Its technical concept is summarized as "Run AI where your data lives", and its capabilities are built at three levels.
The bottom layer is HTAP execution. InnoDB, as the main engine, retains complete transaction semantics and existing operation and maintenance modes. The team's self-developed Rapid in-memory columnar engine, as the secondary engine, undertakes analytical workloads. The two keep synchronized through DML notification, without requiring external ETL jobs. After a query is submitted, the optimizer decides to use the row store or column store based on cost, and the business side is unaware of the engine switch. Rapid adopts vectorized execution internally, and introduces SIMD acceleration on operators such as predicate filtering and aggregation, so that analytical queries and transaction writes can run on the same copy of data.
The middle layer is in-database AI/ML capabilities. Tasks such as classification, regression and anomaly detection can be directly trained and predicted in the database. The tree model side is compatible with mainstream frameworks such as LightGBM and XGBoost. The inference side has a built-in ONNX Runtime, which supports local embedding vector generation and model inference, also supports accessing locally deployed models through Ollama, as well as multiple large model services such as DeepSeek and Tongyi Qwen. The vector retrieval and RAG process form a closed loop in the database, with data and models stored in the same database, avoiding the outflow of features and context. For the approximate nearest neighbor (ANN) index, the team chooses to follow the vector index evolution route of the upstream MySQL community, instead of building a private implementation by itself, to ensure long-term ecological compatibility. All the above capabilities are exposed through standard SQL, and legacy applications do not need to introduce new clients or protocols.
The top layer is the Agent native engine, which is also the core proposition of this project that distinguishes it from external integration solutions: Agent should become a first-class citizen of the database. ShannonBase embeds a lightweight JavaScript engine JerryScript in the server-side process. The Agent executes Javascript code through the internal SQL syntax interface, which naturally reuses the existing permission system, transaction boundary and audit link of the database, without the need to move data to the orchestration framework outside the database. In terms of operation mechanism, Agent adopts the ReAct loop to complete inference and tool calls, and is equipped with a Human-in-the-Loop (HITL) approval state machine: for high-risk actions such as write operations, the execution plan will be stored in the database first and wait for manual confirmation. The approval waiting time is decoupled from the database transaction life cycle to avoid long transaction occupation. Tools and plugins adopt a four-level dispatch mechanism, which takes effect sequentially through session variables, registry, conventional naming and default implementation, making it easy to replace the implementation without modifying the kernel. End users can complete data query and business analysis with one sentence of natural language, and the whole process retains SQL audit and permission verification records that can be traced.
Open-source for user acquisition, subscription for monetization: target users, market space and team progress
ShannonBase adopts the business model of "open-source for user acquisition, monetization via subscription and SaaS": the community version opens core functions for free use, and revenue comes from annual enterprise edition subscriptions, vertical industry solutions, and hosted SaaS on the cloud market. The target users are divided into two categories: the first category is financial, government, public and manufacturing enterprises that hold a large number of MySQL legacy instances and have requirements for data not leaving the database, where in-situ upgrade can avoid migration and compliance costs; the other category is small and medium-sized technology companies that need to quickly add AI capabilities to their business but do not have a dedicated platform team, where a single engine replaces the assembled stack of "database + vector database + feature platform + model gateway", which can significantly reduce deployment and operation and maintenance complexity.
In terms of the market, the project cuts into the intersection of two growing markets: MySQL legacy upgrade and Agent infrastructure. According to industry forecasts, the global Agent infrastructure market is expected to reach tens of billions of US dollars in 2028. If it is estimated on the assumption that about 5% of the tens of millions of global MySQL instances have upgrade demands and the annual subscription for a single instance is thousands of US dollars, the corresponding revenue space is at the level of tens of billions of US dollars. In the domestic market, according to the forecast that the MySQL-related upgrade market will exceed 300 billion RMB in 2028, a 1% penetration rate corresponds to an annual revenue space of about 3 billion RMB. In terms of competition pattern, Oracle MySQL HeatWave is a closed-source solution and deeply bound to OCI, and there is no public cloud region available in Chinese mainland. Although the MySQL Community Edition has supported the VECTOR storage type in its 9.x version, it still has a significant gap from the complete in-database AI and Agent capabilities, and the track has a clear blank area.
In terms of the team, most of the core members of ShannonBase come from database teams of enterprises such as Huawei and Tencent, with more than ten years of kernel R&D experience. They are contributors to open-source communities such as MySQL and openGauss, with capabilities covering the full link from storage engine, query optimizer to executor. In terms of current progress, the initial version has been developed and is undergoing continuous rapid iteration. The official version is expected to be released in October. The code is publicly available in open-source form on GitHub (Shannon-Data/ShannonBase) and open for trial. On the overseas side, the team has released technical content for the Japanese developer community on platforms such as Zenn, Qiita and Note, and launched early market verification. The team will continue to promote market verification in Europe, America and other regions later. The project is currently promoting angel round financing. The funds will be mainly used to expand the database kernel R&D team, invest in the productization and stability verification of the enterprise edition, build a compatibility test system and performance benchmark, and cooperate with early benchmark customers to complete the production environment launch. The remaining part will be used for the continuous operation of the open-source community and developer ecosystem.
MySQL has been serving global enterprises for decades. In the AI era, it should not be directly replaced, but be endowed with new capabilities. The team hopes to use an open-source and open approach to allow all MySQL users to access AI capabilities with low thresholds, and jointly define the AI era of MySQL. The long-term goal is to become the default AI data infrastructure for MySQL users in the AI era, so that data and AI can flow in the same engine.