Salesforce → OpenAI Agents guide

A new guide shows developers how to make an OpenAI agent work inside Salesforce instead of building a separate assistant beside it. (composio.dev) The walkthrough was published by Composio and covers both Python and TypeScript. It uses Composio’s tool router to connect Salesforce to the OpenAI Agents Software Development Kit through the Model Context Protocol, or MCP. (composio.dev) MCP is a standard way for a model to discover tools and call them, like a universal plug that lets one program expose actions to another. The protocol’s server features include tools, prompts, and resources, and its architecture is built around clients talking to servers over a stateful session. (modelcontextprotocol.io, modelcontextprotocol.io) OpenAI describes its Agents Software Development Kit as a framework for building agents that use tools and keep conversation state. Composio’s Salesforce example sets up a session, loads tool access for a user, and lets the agent perform Salesforce operations through natural-language requests. (platform.openai.com, composio.dev) That pattern fits a broader shift in enterprise software: companies want agents to act on data already stored in a system of record, not copy that data into a new interface. Salesforce remains one of the main systems of record for sales, service, and customer data, so tool access inside the platform is the practical step developers need first. (composio.dev, developer.salesforce.com) The guide is also explicit about setup work. Developers need OpenAI and Composio application programming interface keys, must initialize a Composio session for a specific user, and then attach the returned tools to an agent before running prompts against Salesforce. (composio.dev, docs.composio.dev) Salesforce’s own platform rules shape how far these agents can go in production. Salesforce documents organization-level application programming interface limits and exposes a Limits resource so apps can check remaining capacity before they burn through requests. (developer.salesforce.com, developer.salesforce.com) Composio has published similar MCP integrations for other frameworks, including LangChain, which suggests the Salesforce guide is part of a larger push to make one tool layer work across multiple agent stacks. The OpenAI version matters because it gives developers a direct example of wiring an agent framework to a customer relationship management system many companies already run. (composio.dev, composio.dev) The result is less about replacing Salesforce than about giving a model a controlled way to read and act on records that are already there. For teams testing agents in 2026, that is the difference between a demo chatbot and software that can work inside a live business system. (composio.dev, modelcontextprotocol.io)