Mellea

Mellea's core bet is that the LLM call should look like any other function in your codebase: a typed signature, a contract, and a test. The @generative decorator makes the docstring the prompt and the type hints the output schema, which removes two of the most common sources of drift in generative code — prompt templates stored separately from the code that calls them, and parsers that silently disagree with the schema. The declared-requirements model goes further: tone, length, content rules and custom predicates are checked before a response leaves the call, with automatic retry and repair on failure. Where it genuinely separates from the field is constrained decoding. On Ollama, vLLM, and HuggingFace the library constrains generation grammatically at the token level, so malformed structure can't be produced in the first place. That is a materially stronger guarantee than validate-and-retry, and the homepage is explicit that this is the distinction from Instructor and PydanticAI. The honest corollary, stated in the project's own limitations, is that backends without constrained-decoding support — OpenAI, Bedrock, Watsonx, LiteLLM — fall back to validation plus repair, so the strength of the guarantee depends on where you deploy. Confidence tuning is well designed for production rather than demo use: majority voting and best-of-n are a single parameter change, not an architecture change. The agent surface matured quickly — v0.6.0 (May 19, 2026) brought streaming with per-chunk validation, typed events, an OpenTelemetry bridge, MCP tools, context compaction, a sandboxed code interpreter and a shell tool; v0.8.0 (Sep 24, 2026) added audio input, retrievable files from the sandbox, rebuilt streaming as a plain async iterator and cut CLI startup from 3.2s to 0.1s. If you use streaming or telemetry, v0.8.0 carries breaking changes worth reading before you upgrade. Granite Guardian integration gives you harmful-output, hallucination and jailbreak checks without a sidecar service. The trade-offs are real and mostly structural. It is Python-only with no port to other stacks, and it expects comfort with type annotations and Pydantic-style models. There is no visual builder and no hosted dashboard — you own deployment, testing and failure handling. The project is young (v0.6 in May 2026, v0.7 in July 2026, v0.8 in September 2026), so community size and documentation depth trail established frameworks. The vision work is a good example of where it's headed: the July 28, 2026 blog post on getting typed objects out of images with Granite Vision 4.1 via the format= parameter and ImageBlock shows structured extraction extending past text. For a team already writing tests, Mellea slots into CI naturally; for a team looking for a canvas or a managed endpoint, it is the wrong shape entirely.