python-sdk

Install the package with pip install inferencesh, create a client with inference(api_key="inf_..."), then call client.run with the app ID and input dictionary. The result contains the status and output data.

Use the tool builder API to define tools with tool(), app_tool(), agent_tool(), or webhook_tool(), specifying parameters with types like string, number, or enum_of. Pass the tools list to client.agent along with a core_app model reference and system prompt.

Yes, the SDK provides an async_inference client for asyncio-based code, supporting await client.run and async iteration over streaming updates. The sync client also supports stream=True to receive real-time progress updates and logs.

Local file paths in the input dictionary are uploaded automatically. For manual control, call client.upload_file with a path, bytes, or file object, plus UploadFileOptions for filename, content type, and public access.

Sessions keep workers warm between requests by passing session="new" to create one and reusing the returned session_id. Use them for model loading, browser automation, or multi-turn conversations to avoid cold starts, with timeouts from 1 to 3600 seconds.

Call .require_approval() when building a tool, then check call.requires_approval in your on_tool_call handler. Prompt the user for confirmation and submit the result or a denial via agent.submit_tool_result.