OpenAI Agents SDK
An SDK (software development kit) is a set of ready code and tools for a programmer. OpenAI released the Agents SDK on 11 March 2025 together with the Responses API, which has built-in file search and web search.
It rests on three concepts. An agent is a model with instructions and tools. A handoff lets one agent pass work to another or call it as a tool. Guardrails check what comes into the agent and what it answers.
Around them are sessions (the memory of a conversation), tracing (a record of every agent step for debugging), connections to MCP servers and human approval of actions.
The code is open under the MIT license: you can use and change it in commercial projects. It installs with one command, pip install openai-agents, and the model key goes in the OPENAI_API_KEY variable.
The kit works not only with OpenAI models. The documentation describes connecting any vendor with an OpenAI-compatible API through its own address and key, plus the LiteLLM and Any-LLM adapters, which are still in beta.
By default, tracing sends records to OpenAI. Without an OpenAI key, the documentation advises turning it off with the set_tracing_disabled() function or connecting your own record handler.
Claude Agent SDK
The Claude Agent SDK grew out of Claude Code, Anthropic’s agent for programmers. On 29 September 2025 Anthropic renamed Claude Code SDK to Claude Agent SDK: a loop of “gather context, take an action, check the result, repeat” and connections to outside services through MCP. MCP is an open standard for connecting AI to data and services.
According to the documentation, the kit gives the same tools, agent loop and context management as Claude Code, in Python and TypeScript. The built-in tools read and edit files, run commands and search the web.
On top come hooks (your own code at key points of the agent’s work), subagents (helper agents for separate subtasks), permissions (which tools run on their own and which wait for approval) and sessions that you can resume later.
Requirements: Python 3.10 or Node.js 18 or newer. Install it with pip install claude-agent-sdk or npm install @anthropic-ai/claude-agent-sdk. Access to the model is by the ANTHROPIC_API_KEY key or through the Amazon Bedrock, Google Cloud and Microsoft Foundry clouds.
The terms differ from OpenAI’s: Anthropic’s commercial terms govern the use of the kit, including when you build a product on it for your own clients. Third-party products may not log in through a claude.ai account without Anthropic’s permission; only an API key is allowed.
How they differ
The OpenAI kit is a light framework: you write the tools, and you decide where the agent runs. Its main idea is passing tasks between agents and checks on input and output.
The Claude Agent SDK is the ready Claude Code agent as a library. It works with files and commands from the first run, so it fits tasks where the agent needs a computer: sort documents in a folder, prepare a report, fix code.
Models. According to its documentation, the OpenAI kit connects other vendors and local models with a compatible API. In its documentation, the Claude Agent SDK works with Claude models through the Anthropic API or through intermediary clouds.
License. The OpenAI kit’s code is open under MIT. The Claude Agent SDK is used under Anthropic’s commercial terms.
Both share the agent loop, tools, MCP connections and step logging. The third path for a developer, the open LangChain and LangGraph libraries, is not tied to one model vendor.
Availability in Russia and open models
Russia is not on the list of countries where the OpenAI API works. The OpenAI documentation warns that access from countries outside the list may end with a blocked account.
Anthropic is the same: Russia is not in its supported regions policy of 12 May 2026, and anything not on the list is not supported.
There is also the data question. In the basic setup, both kits send requests to the model vendor’s servers abroad. If the requests contain clients’ personal data, under 152-FZ the operator must file a notice with Roskomnadzor, the Russian data regulator, before a cross-border transfer, and the data of Russian citizens is collected and stored in databases in Russia. How we solve this is described on the page about data security.
The practical path is open models on your own computer or server through Ollama. Ollama serves models through an OpenAI-compatible API at http://localhost:11434/v1/. The client requires an API key, but Ollama does not check it. You download the model first with the ollama pull command.
For the OpenAI kit this is a standard path from the documentation: your own client with the Ollama address through set_default_openai_client or the OpenAIChatCompletionsModel model, with tracing to OpenAI turned off. The code is under MIT, and the model and the data stay with you.
Ollama also answers in the Anthropic API format: its documentation describes connecting Claude Code to local models through the ANTHROPIC_BASE_URL variable. We found no such instructions for the Claude Agent SDK from either Anthropic or Ollama, and Ollama does not support some features, such as request caching.
How agents on a local model work and which server they need, we covered separately.