Nola
nola.sh- Category
- Developer Tools
- Rank
- No. 3342Tools index
Previous survey · No. 3209 ·
- Type
- TOOL
- Use case
- Coding
- Interfaces
- CLI
- GitHub
- 7 stars
- Latest release
- v0.1.14
- Date
About
Nola is a TypeScript superset (.tsi files) that treats LLM inference as a native language construct rather than an SDK call, using 'infer' and 'ask' keywords so typed values and function-calling are checked by the compiler instead of hand-written schemas. It compiles down to plain TypeScript and includes a CLI, editor support, and observability tracing for inference calls.
What it does
Nola adds two keywords to TypeScript so that a request to a language model reads like an await. You write a TypeScript interface, then ask the model for a value of that type in a single expression. The compiler turns the file into ordinary TypeScript, works out the response format from the type, and checks the model's reply against it when the program runs. A reply that does not fit gets one correction attempt before the call fails with an error. Plain TypeScript files can import these functions and await them like any other async code.
Why it's ranked here
The design removes the second copy of the data model that most typed model-calling code carries. The project's own six-way comparison shows the domain type declared once in Nola and twice in LangChain.js, Ax, the Vercel AI SDK and the plain OpenAI SDK, with no code generation step, against BAML's fourteen generated files. It is Apache 2.0 licensed, ships editor tooling and plugins for seven bundlers, and supports OpenAI, Anthropic, Google and any compatible endpoint. The caveat is that the comparison was written by the authors and measures code organisation, not model quality, latency or cost.
What's good
Context is handled precisely. A parameter marked with a leading dot is written into the prompt for every request in that function, while an ordinary parameter only shows its name and type to the model, never its value. Classification questions can return a probability for each answer plus a confidence figure instead of a single guess. Calls are deferred: nothing reaches a provider until the result is awaited, and each one resolves at most once. A deterministic mock provider, record and replay ledgers, and a setting that forces every request to the mock in CI make tests runnable with no key and no network.
Tradeoffs
Adopting it means adopting a new file type. The documentation states the cost plainly: you need the compiler, a loader or bundler plugin, and editor support, because the standard TypeScript compiler never sees these files. That is more tooling than a library import. Node 22.18 or newer is required. The hosted platform that powers the no-key first run is a service with trial limits, so production use means configuring your own provider keys. The comparison explicitly leaves out streaming and React bindings, which are strengths of the Vercel AI SDK, so interactive chat interfaces are not the case it argues for.
How to use it well
It fits a TypeScript backend that pulls structured data out of text, routes tickets, or classifies messages, and whose team already keeps its data model in interfaces. Start with the scaffolder and the offline templates, then switch on a real provider in the config file. Use named models for cheap drafts and careful final passes, and force the mock provider in CI so tests stay deterministic. Keep the compiler as a development dependency, since only the runtime and provider packages ship. Look elsewhere for streaming chat UIs or prompt optimisation, which the comparison pages leave to other tools.
Technical notes+
The repository is an npm workspaces monorepo (package.json, workspaces packages/* and examples/*, Vitest for tests, Biome for lint). packages/compiler/src/index.ts exports compileNola, compileView and the schema helpers collectTypeRegistry and collectExportedTypeNames, which is where types become wire schemas at compile time. packages/runtime/src/ask/index.ts exposes ask, validate, wireSchema, runPipeline and fingerprintRequest from @nola-lang/core; the fingerprint keys replay entries, and docs-site/config/providers.mdx notes the inference profile is part of it. packages/providers/src/index.ts exports the openai, anthropic, google, typesafe and mockProvider factories plus withRetry, fallback, roundRobin, record and replay. Routing precedence is forceModel, then ask with or withModel, then model.default; an unknown name raises NOLA3004.
Observed
- License
- Apache License 2.0
- Language
- TypeScript superset compiled to plain TypeScript
- Runtime requirement
- Node 22.18 or newer
- Interfaces
- CLI (init, build, run, check, declarations, skill), Node loader, runtime library
- Install surface
- npm create nola scaffolder; nola-lang as a dev dependency
- Bundler plugins
- Vite, webpack, Rollup, Rolldown, esbuild, Rspack, Next.js
- Editor support
- tsserver plugin, language server, VS Code extension
- Providers
- OpenAI, Anthropic, Google, typesafe.ai, OpenAI-compatible endpoints, hosted platform, mock
- Test runner
- Vitest, declared in the root package.json
Read from README.md, package.json, LICENSE, docs-site/index.mdx, docs-site/compare/index.mdx, examples/extract-person/src/main.ts, packages/compiler/src/index.ts, packages/runtime/src/ask/index.ts, packages/providers/src/index.ts, docs-site/config/providers.mdx, comparisons/inbox-triage/README.md.
What it can do
Compile .tsi TypeScript superset files to plain TypeScript
.tsi source file → TypeScript code
Request typed model outputs using the 'infer' keyword
Typed inference request in code → Typed model output
Trigger native function calls using the 'ask' keyword
Function call request in code → Function call result
Trace inference calls for observability
Inference call execution → Trace data
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.