OpenAI Adds Remote MCP Support and New Built-In Tools to Responses API

OpenAI has added support for remote Model Context Protocol (MCP) servers to its Responses API, its core API primitive for building agentic applications, alongside new built-in tools for image generation, Code Interpreter, and an upgraded file search. The tools work across the GPT-4o series, GPT-4.1 series, and the OpenAI o-series reasoning models, the company announced.

The stakes are significant for the developer ecosystem. Since OpenAI released the Responses API in March 2025 with web search, file search, and computer use tools, hundreds of thousands of developers have used it to process trillions of tokens across the company's models, according to OpenAI. Customers already building on the API include Zencoder's coding agent, Revi's market intelligence agent for private equity and investment banking, and MagicSchool AI's education assistant — all of which use web search to pull current information into their apps.

Reasoning models now call tools inside their thinking

The most consequential change affects the reasoning models. o3 and o4-mini can now call tools and functions directly within their chain-of-thought in the Responses API, which OpenAI says produces answers that are more contextually rich and relevant. By calling multiple tools while reasoning, the models achieve significantly higher tool calling performance on industry-standard benchmarks like Humanity's Last Exam, the company reports.

Using o3 and o4-mini with the Responses API also preserves reasoning tokens across requests and tool calls, improving model intelligence while reducing cost and latency for developers.

Remote MCP servers, with OpenAI joining the steering committee

MCP is an open protocol that standardizes how applications provide context to LLMs. With the new support in the Responses API — building on the earlier release of MCP support in the Agents SDK — developers can connect OpenAI's models to tools hosted on any MCP server with just a few lines of code.

Popular remote MCP servers already available include Cloudflare, HubSpot, Intercom, PayPal, Plaid, Shopify, Stripe, Square, Twilio, and Zapier. OpenAI expects the ecosystem to grow quickly in the coming months, making it easier for developers to build agents that connect to the tools and data sources their users already rely on.

The company has also joined the steering committee for MCP. That move puts OpenAI in a formal governance role for a standard that originated outside the company — a signal of where interoperability for agentic AI is heading, and of how much weight the protocol now carries across payments, commerce, and communications infrastructure.

Image generation, Code Interpreter, and file search

With built-in tools in the Responses API, developers can create more capable agents with a single API call. Three tools joined the lineup today:

• Image generation. Developers can now access OpenAI's latest image generation model, gpt-image-1, as a tool within the Responses API, in addition to the Images API. The tool supports real-time streaming — developers see previews of the image as it generates — and multi-turn edits, allowing granular, step-by-step refinement of images.
• Code Interpreter. The tool, useful for data analysis, solving complex math and coding problems, and manipulating images (OpenAI cites its "thinking with images" capability), is now available in the Responses API. The ability of o3 and o4-mini to use Code Interpreter within their chain-of-thought has improved performance across several benchmarks, including Humanity's Last Exam.
• File search. The tool is now available on OpenAI's reasoning models. It pulls relevant chunks of developers' documents into the model's context based on the user query. Updates add the ability to search across multiple vector stores and support attribute filtering with arrays.

Reliability, visibility, and privacy features

OpenAI also shipped three features aimed at enterprises. Background mode handles long-running tasks asynchronously: reasoning models can take several minutes to solve complex problems, as seen in agentic products like Codex, deep research, and Operator. Developers can now build similar experiences on models like o3 without worrying about timeouts or connectivity issues, either polling for completion or streaming events when their application needs to catch up.

Reasoning summaries give developers concise, natural-language summaries of the model's internal chain-of-thought, similar to what ChatGPT shows. OpenAI says this makes it easier to debug, audit, and build better end-user experiences. The summaries are available at no additional cost.

For privacy-sensitive deployments, customers eligible for Zero Data Retention (ZDR) can now reuse reasoning items across API requests without any reasoning items being stored on OpenAI's servers. For o3 and o4-mini, reusing reasoning items between function calls boosts intelligence, reduces token usage, and increases cache hit rates, lowering costs and latency.

Pricing and availability

All of the new tools and features are available now in the Responses API, supported across the GPT-4o series, GPT-1.1-series-equivalent GPT-4.1 models, and the o-series reasoning models (o1, o3, o3-mini, and o4-mini). Image generation is supported only on o3 among the reasoning models.

Pricing for existing tools is unchanged. Image generation costs $5.00 per 1M text input tokens, $10.00 per 1M image input tokens, and $40.00 per 1M image output tokens, with 75% off cached input tokens. Code Interpreter costs $0.03 per container. File search costs $0.10 per GB of vector storage per day and $2.50 per 1,000 tool calls. Calling remote MCP servers carries no additional tool fee — developers are simply billed for output tokens from the API.

With MCP steering committee membership, in-context tool calling for its reasoning models, and encrypted reasoning for ZDR customers, OpenAI is positioning the Responses API as the default substrate for enterprise agents — and the coming growth of the remote MCP server ecosystem will determine how much of the tooling layer it can realistically consolidate.