Mirascope

Mirascope is an open-source toolkit from Mirascope, Inc. for building LLM-powered applications in Python, with a TypeScript implementation also present in the repository. It provides one interface for calling frontier models from providers such as OpenAI, Anthropic, and Google, and describes itself as "The LLM Anti-Framework". The latest GitHub release listed is v2.5.0, published 2026-06-24.

What It Is

Mirascope is a developer library for calling LLMs, defining tools, parsing structured output, and building agents. Developers decorate ordinary functions to turn them into model calls or tools, rather than adopting a heavyweight framework. The documentation lists provider-agnostic APIs, built-in observability, and production features such as tools, structured outputs, and error handling.

How the Workflow Looks

A function decorated with @llm.call takes a model identifier such as an OpenAI or Anthropic model, and can accept tools and a Pydantic model as the output format. Responses can be parsed into typed objects. Agent loops are written explicitly: while the response contains tool calls, the code executes the tools and resumes the response with their outputs. The docs also cover messages, streaming, and async usage.

Observability and Versioning

The ops module adds OpenTelemetry-based tracing, sessions, and versioning. Decorating a function with @ops.version tracks changes to prompts and records traces, and @ops.trace captures nested execution such as a full agent loop. The homepage example shows traces with version, token counts, and cost per call. A Mirascope Cloud offering is also linked from the site.

Setup Path

Install with uv add "mirascope[all]" or pip, then set a provider API key as an environment variable. The repository is MIT licensed and organized as a monorepo with Python, TypeScript, website, and docs directories.