OpenAI Knowledge

OpenAI Knowledge is an official skill from the openai/openai-agents-python repository. It defines a controlled research path for questions about the OpenAI API and platform features. The primary source is the .agents/skills/openai-knowledge directory in the official Agents SDK repository. according to the provider, an agent should search the OpenAI Developer Documentation through the official Docs MCP server and then fetch the relevant documentation pages as exact Markdown content. The answer should be based on that retrieved text rather than guessed or outdated model knowledge.

Purpose and workflow

The skill is a procedure for documentation-grounded technical answers. It is especially suitable for questions about the Responses API, tools, streaming, the Realtime API, authentication, models, rate limits, and MCP. The first step is checking whether the OpenAI Developer Documentation MCP tools are available in the selected environment. The agent then searches for a relevant official page and fetches the specific result. This search-and-fetch sequence makes the evidence traceable and reduces the risk that an answer comes from an old training snapshot or a search snippet.

For questions about endpoints, parameters, response fields, or schemas, the skill provides for additional OpenAPI tools. According to the provider, an OpenAPI specification and an endpoint list can be used when those tools are available at runtime. The instructions explicitly require the agent not to invent flags, field names, defaults, or limits. This matters in day-to-day development because small differences in JSON fields, authentication, or version behavior can produce broken integrations.

Team use

Teams can use OpenAI Knowledge as a reusable research foundation before implementation, code review, or a technical decision. One person can formulate a concrete question while the agent locates the responsible official page, fetches its content, and checks the claim against the source. This supports traceable decisions about API selection, tool configuration, model migration, MCP integration, and troubleshooting. Responsibility remains with the relevant specialists, who must evaluate the documentation, the review time, and the impact on their architecture.

The skill is not an API client, a standalone MCP server, or a runtime environment. It provides no credentials and stores no tokens, secrets, or project-specific settings. If the Docs MCP tools are not configured, the source only explains how an authorized operator can enable them in the operator's Codex environment; the skill does not configure them by itself. Availability therefore depends on the selected harness and its MCP configuration. The official source explicitly names Codex, while compatible environments may apply the procedure in an equivalent way.

Security and limits

Source-oriented research improves freshness but does not guarantee a correct or approved solution. A successful documentation fetch proves neither that an account may use a feature nor that a region, organization, model entitlement, or local policy is compatible. Test examples in a controlled environment before production use and do not send confidential inputs to unapproved tools. Documentation is evidence, but it is not automatic approval for changes to production systems.

This description relies on the official skill file, the OpenAI Agents SDK repository, and the official documentation for the Docs MCP server. Provider statements are marked as such. Data transmission depends on the selected agent, enabled MCP tools, and model route. A local agent does not automatically mean that processing by the model provider remains local. The skill is therefore useful for verifiable research, but it does not replace security review, privacy review, access-control checks, or change approval. Recheck current documentation pages and the repository license before redistribution.