Build Hour: Built-In Tools

Built-in tools let large language models search the web, query private files, and call external services—without developers writing the usual glue code to run those actions. Instead of manually executing function calls and feeding results back into the model, the tools run on OpenAI infrastructure and the outcomes are automatically inserted into the conversation so the model can produce a final answer in the same flow. That shift matters because it removes a major source of engineering overhead (and state-management complexity) when building agentic apps that need live information or access to business systems.

The session breaks “built-in tools” into two parts. “Tools” are required when LLMs must interact with data or take actions—LLMs can generate text, but they can’t natively search internal databases, retrieve current events, or operate on third-party platforms. “Built-in” means developers don’t have to code the tool plumbing themselves; they can enable a set of hosted capabilities and grant the model access. The team contrasts this with traditional function calling: function calling requires three explicit steps—declare functions, execute them in your code, then return results to the model. Built-in tools collapse that middle execution step by running the tool automatically and appending results to the context.

As of the time of the demo, six tools are available. Web search addresses the models’ knowledge cutoff (described as May 2024 for current models), giving real-time access to public information via OpenAI’s own services and index. File search provides retrieval over an uploaded knowledge base using the RAG pattern (retrieval-augmented generation), handling preprocessing like chunking and embedding as well as retrieval and reranking so developers don’t need to build a full retrieval pipeline. The MCP tool (Model Context Protocol) is positioned as a gateway to hundreds of remote tools: by connecting to an MCP server, models can access provider-specific capabilities exposed by that server. Code interpreter executes Python on OpenAI infrastructure for deterministic computation, data analysis, and chart generation, and it can also work with files uploaded into the tool environment.

A live walkthrough uses the OpenAI Playground as an experimentation lab. File search is demonstrated by uploading two PDFs and asking a question; the model returns bullet-point answers with citations to the specific files it used. Web search is demonstrated with a current-events query (including a note that web search can “spiral” when it keeps searching, so prompts should instruct it to search once). MCP is demonstrated with Shopify: the model fetches the tools available on a specific store, asks for approval when configured to do so, and then returns product results without developers manually wiring Shopify APIs. Code interpreter is shown both alone and in combination with web search, including generating a comparison chart from historical weather data.

The second half turns the demos into an application pattern. A sample dashboard app is built by gradually adding tools—Stripe via an MCP server, web search, and code interpreter—so users can ask questions like “What’s my Stripe balance?” and get both answers and visual components. The app uses a custom “generate component” function with structured outputs to render cards, tables, and charts, while hosted tools handle data retrieval and computation. The practical takeaway: built-in tools reduce time-to-prototype, offload retrieval and execution complexity, and support multi-turn tool use where the model can chain actions to solve multi-step tasks.

A customer spotlight from Hebia ties the approach to financial and legal workflows. Hebia uses web search to overcome knowledge cutoffs and to support an explore/exploit research strategy, then deepens analysis using sources like SEC filings and private datasets. In its products, web search is used at scale (e.g., adding columns across thousands of companies) and as part of long-running research plans that can span up to an hour, with MCP used selectively when off-the-shelf servers don’t meet enterprise indexing needs.