Agently

Agently is an open-source Python framework for engineering teams building AI service backends on their own infrastructure. It focuses on turning model responses into validated data and controlled actions that business applications can use. The framework runs in your application; requests go to the model provider you connect, including external services.

Its output schemas define typed fields and required values, with validation and retries to handle incomplete responses. Structured streaming lets an interface update individual fields before generation finishes. Those partial values remain provisional until final validation, a useful distinction for applications that act on model output.

Agents can call local functions, connect to MCP servers, and execute code in sandboxes. Permission boundaries control actions, while execution records capture results and references to generated artifacts. Reusable runtime Skills supply task guidance, with the host application retaining control over permitted actions.

TriggerFlow handles branching, parallel work, event-driven inputs, and pause or resume. Persistence and restart-safe execution support longer processes, including workflows with human review. Dynamic Task validates and runs task graphs supplied by applications or generated by models.

Agently supports OpenAI-compatible, OpenAI Responses-compatible, and Anthropic-compatible APIs without requiring a rewrite of business logic. FastAPI integration exposes agents as service endpoints with streaming responses. Agently-DevTools provides a local console for logs, evaluation, and run inspection. The framework uses the Apache 2.0 license.