Tangent sits between your spec and your coding agent — constraining AI generation through opinionated templates while staying intelligent exactly where it needs to be.
pip install tangent-cliPredictable, reviewable — but too rigid to express real logic.
Too rigidFlexible, fast — but unauditable, non-reproducible, and expensive.
Too looseFour primitives that turn AI code generation into a repeatable engineering process.
Write templates that embed LLM calls only where static output can't express the logic. Everything else is deterministic by construction.
Tangent builds a dependency graph across all LLM calls so outputs are generated in the correct order, every time.
Every result is hashed and stored. Only inputs that actually changed trigger a new LLM call — the rest is served from the cache.
The lockfile records exactly what was generated, from what inputs, with what model. Commit it. Review it. Diff it like any other source artifact.
reduction in AI inference costs on repeat runs — because unchanged inputs never re-invoke the model.
The same tool that helps a solo hacker ship fast also satisfies an enterprise compliance team.
Ship fast with AI and actually be able to maintain what you shipped. Every generated file conforms to the architectural patterns you reviewed once — not whatever the agent felt like that day.
Replace unauditable AI output with traceable, reproducible generation. Review architectural decisions once at the template level. Every subsequent generation is consistent by construction.
The lockfile is a tamper-evident audit trail. Planned policy enforcement constrains filesystem, network, and dependency usage — containing the compliance risk that unconstrained generative code creates.
Open source. Designed to fit into your existing workflow.
Get started on GitHub