TypeLLM vs Marvin

These libraries attack the same problem — typed output from an LLM — from opposite ends. Marvin is a general-purpose framework: decorate a function with @ai_fn and it becomes an AI call, @ai_classifier handles text classification, and structured extraction rides on Pydantic models. Around that core it ships product plumbing you'd otherwise build: agent loops with tool calling, SSE streaming, async-first API, rate limiting, retries, concurrency control, a CLI monitor and a SQLite state store. It talks to OpenAI and Anthropic, so embedding generation and chat are on the menu. TypeLLM is narrower and more opinionated: you declare each output field as a JSON Schema entry with a type (string, integer, number, boolean, enum) and plain-English instructions, and the library constrains generation so values are schema-conformant rather than parseable-by-luck. Two capabilities stand out. First, dependency-graph ordering via depends_on lets later fields consume earlier fields, with fields otherwise run in parallel. Second, and rarer, permutation-invariant decision probabilities for constrained choices make it usable for classification pipelines that need calibrated preference rather than a single label. Thinking mode, image input, multi-thread clients, per-call timeout/cancel/seed and balanced permutation averaging round it out. The trade-off is ecosystem: Marvin plugs into closed APIs and your existing app; TypeLLM is a Python client against an SGLang HTTP endpoint, designed for open autoregressive models you serve yourself.

Marvin is free and open-source — there is no license cost, seat count or usage meter. Your real bill is whatever OpenAI or Anthropic charges for the calls your decorated functions make, plus your own compute and time. That makes budgeting straightforward and low-risk to trial: install it, run a test suite against a cheap model, and you know your unit economics. The hidden costs are operational — you own retries, rate-limit tuning, state and any orchestration you grow into, which is precisely the work Marvin partially absorbs with its built-in retry, rate-limiting, concurrency and SQLite state features. TypeLLM's pricing_type is 'contact', so there is no public price to compare; you negotiate. Its economics are different in kind: the selling point is minimal compute cost versus generating and then parsing free text, and it targets self-hosted open models, so the dominant line item is your GPU capacity, not a per-token API invoice. Add engineering time for standing up and maintaining SGLang, plus the client plumbing that 0.1.7–0.2.3 kept reshaping. For a solo developer, Marvin is the near-zero-cost path; for a team already paying for GPUs, TypeLLM can reduce tokens per extraction, but you must obtain a quote and amortize the infra you're already carrying.