Gemini Deep Research gets MCP integration, collaborative planning, and two new versions

On April 21, 2026, Google published a significant expansion of the Deep Research agent in the Gemini API changelog. Developers now have access to two new model versions, streaming responses, integration with MCP (Model Context Protocol) servers, and collaborative research planning.

What happened?

Google released two new models in the Gemini API: deep-research-preview-04-2026 and deep-research-max-preview-04-2026. The first is, according to the official description, “designed for speed and efficiency, ideal for streaming to client interfaces,” while the second targets “maximum comprehensiveness for automated context gathering and synthesis.”

Alongside the new models come four key capabilities: streaming responses (partial results arrive while the agent is still working), collaborative planning (the agent first proposes a plan that the user can edit before execution), visualizations (the agent can generate graphs and diagrams as part of results), and MCP server integration with File Search support.

MCP (Model Context Protocol) is an open protocol that Anthropic first standardized, and Google is now building support for it deep into the Gemini platform. Practically speaking, this means the Deep Research agent can access the user’s own databases, CRM systems, internal wikis, or any other MCP-compatible source during research.

Why does this matter?

Deep Research as an agent category has been a key battleground between major AI labs for a year — ChatGPT Deep Research, Perplexity, and Anthropic Research have all introduced similar agents. Google’s advantage had previously been in synthesis quality, but they lagged in integration flexibility.

MCP integration changes that equation. Instead of Deep Research drawing exclusively from the public web, enterprise clients can now connect it to their own data sources through a standard protocol — without writing custom connectors for each source. For companies, this opens scenarios like “research our ticketing system and synthesize the main causes of reported errors in the past 90 days” without building a custom RAG stack.

Collaborative planning addresses the previous complaint about agents — that they “go in the wrong direction” and the user loses 5–10 minutes of waiting time. By showing the plan before execution, the agent becomes a predictable tool rather than a black box.

What’s next?

Both versions are currently in preview status, meaning Google retains the right to change the API before a stable release. Developers already using the Gemini API can try them immediately, and full documentation is available on the official Deep Research page of Google’s AI platform.

For the local development ecosystem, the key question is pricing — the changelog does not include billing details, but the max-preview variant is expected to be significantly more expensive than the preview variant due to longer token counts and deeper synthesis. The practical recommendation is to use the preview model for interactive applications where streaming is important, and the max model for overnight batch jobs where maximum report quality is required.

Together with yesterday’s Google ReasoningBank work, a clear pattern is visible — Google is positioning the Gemini ecosystem as a platform for agents that work with external tools and accumulate knowledge, not just as a chatbot.