MongoDB adds AI retrieval tools to Atlas platform

MongoDB has added AI data retrieval features to its Atlas database platform and expanded integrations with developer tools including ChatGPT, Claude, Codex, Grok Build and Devin. The update brings Voyage AI models into the database and introduces a managed server for connecting coding agents to live operational data.

The changes are intended to let developers query, inspect and update MongoDB data from within the tools they already use to build software, while reducing the work involved in connecting AI agents to production data. Atlas Managed MCP Server is hosted within Atlas and uses the same credentials and access controls as the database service itself.

MongoDB's latest product push combines two AI infrastructure layers that have often sat apart: retrieval models for finding relevant context and direct links between coding assistants and live application data. By placing both inside Atlas, MongoDB is making the case that operational databases should also serve as the memory and retrieval layer for AI applications and software agents.

The new integrations span consumer and developer-facing AI products. Users can ask questions about data in plain language in ChatGPT, Claude and Grok, while coding agents such as Claude Code, Codex, Grok Build and Devin can query and modify data during development. Developers can also view live MongoDB data while generating applications inside integrated development environments such as Cursor.

In practical terms, the connectors allow AI tools to list collections and indexes, query and aggregate data, and inspect schemas. Users with the right permissions can also create collections and manage indexes. The setup requires marketplace authorisation rather than manual configuration of connection strings or extra infrastructure.

The managed MCP server is aimed at teams running agents in production. MCP, short for Model Context Protocol, has emerged as one way to link models and agents to external tools and data sources. MongoDB said its existing MCP server already records more than 30,000 installs a week, and the new managed version removes the need for customers to install, operate or upgrade that connection themselves.

"The AI tools teams reach for keep changing, so our approach is to make sure MongoDB is present in all of them, whether a team is working in Claude or Codex, or running an agent in production. More of that building is now done by agents, and neither the agent nor the developer has to stop and set up a connection, so applications come together faster," said Pablo Stern-Plaza, chief product officer, AI and emerging products, MongoDB.

OpenAI and Cognition backed the integrations with customer-style endorsements focused on access to live application context.

"Developers want their AI tools to connect with the context and systems they already rely on," said Vibhor Chhabra, product lead for ChatGPT ecosystem at OpenAI. "MongoDB's plugin in ChatGPT makes it easier to access and work with live application data, helping developers move faster while staying grounded in the context of their applications."

"We're in the golden age of software engineering. The scope of what one engineer can build has exploded, and the unlock is agents working with real context," said Russell Kaplan, president at Cognition, the company behind Devin. "By connecting Devin to MongoDB Atlas, engineers can hand off well-scoped tasks knowing Devin is working from live application data, not stale assumptions, and spend their own time on the harder problems."

Retrieval push

Alongside the tool integrations, MongoDB has brought Voyage AI's embedding and reranking models into Atlas. The additions include Automated Embeddings in Atlas, an Atlas Embedding and Reranking API, a code-focused model called voyage-code-4, and vector search in Atlas Stream Processing for data in motion.

The aim is to address a common problem in AI systems: retrieval quality depends on current, relevant data, but many organisations still move information out of operational systems into separate vector databases and search layers. That can create synchronisation overhead, extra infrastructure and the risk of agents acting on stale information.

MongoDB's approach is to embed and index documents within the operational database as data is written or changed. Users can choose a Voyage AI embedding model for a search index, after which Atlas handles embedding for new and updated documents automatically.

The standalone Atlas Embedding and Reranking API extends those models to applications running outside MongoDB. The code-specific voyage-code-4 model is designed for retrieval tasks involving software repositories and coding agents, where general-purpose embedding models can struggle to match code and intent accurately.

Customer examples

Financial Times and legal AI platform Eve were cited as early users of the retrieval features.

"Our job is to make the FT's journalism fast and easy to reach, however our readers come to it. With Automated Embedding and Voyage AI models on Atlas, we've improved retrieval accuracy while keeping costs in check across more than 100,000 searches a day, and being able to test and balance models lets us tune quality against cost as we go. With less infrastructure to run, the team can spend more of its time on the reading experience our subscribers rely on," said Elitsa Pavlova, principal engineer CM platform at Financial Times.

"In legal AI, retrieval quality is foundational-the right evidence has to surface at the right moment. MongoDB's Atlas Embedding and Reranking API gives us a promising way to improve relevance directly in the RAG layer, while simplifying the infrastructure needed to build and evolve these experiences," said Urvesh Patel, staff AI engineer at Eve.

MongoDB's chief technology officer framed the launch as a response to fragmented AI stacks that combine an operational database, vector store, search engine and model providers from different vendors.

"Too many organizations are running AI in production with an operational database, a vector store, a search engine, and embedding and reranking models, all from different vendors, bolted together instead of built for it," said Jim Scharf, chief technology officer at MongoDB. "That's where stale data and errors creep in, and it's usually where teams spend their time babysitting instead of building. Agents raise the bar. They need to retrieve live context continuously and cannot wait on overnight batch jobs. MongoDB was built as an operational platform from the start, so retrieval and memory run on the same live data, nothing to sync, and agents act on what's happening instantly."