A language for AI workflows
Stop reinventing
the AI runtime.
Voss makes confidence checking, token budgets, prompt construction, semantic routing, and agent lifecycle first-class language constructs — not boilerplate you write in every project.
From the cloned repo. PyPI release with v1.
Five things every AI app reinvents.
Plus one.
Voss makes them language constructs — checked by the compiler, owned by the runtime, written once.
Confidence as a type
Calls to a model return a value with a confidence score. Gates like `if intent @ 0.80` are checked by the compiler — no more silently trusting low-confidence output.
Token budgets, not prompt math
`ContextScope(token_budget=3000)` is a language construct. Voss handles compression, eviction, and budget enforcement so your code stays declarative.
Semantic routing, compile-time
Embedding-based `match` cases are computed at build time. Route by meaning without paying for an embedding call on every request.
Agent concurrency primitives
Spawn researchers in parallel, gather their results with a timeout, fall back gracefully on budget exhaustion. Like goroutines for agents.
Anthropic, OpenAI, Ollama
One runtime, swappable providers. Your Voss program doesn't care which model is behind it — switch with a config flag, not a rewrite.
Readable Python out
Voss compiles to debuggable Python you can read, diff, and own. No black-box runtime, no magic — your generated code is shippable on its own.
Three patterns. One CLI.
Run any example from the examples/ folder. The compiler accepts .voss source; the runtime works with raw Python today.
from voss_runtime import ContextScope, ProbableValue
async def classify_intent(user_input: str) -> str:
async with ContextScope(token_budget=1000) as ctx:
await ctx.add(f"Classify the intent: {user_input}")
intent: ProbableValue = await ctx.ask(
"Return only the intent label.",
return_type=ProbableValue,
)
if intent @ 0.80:
return intent.value
return "unknown"The CLI.
One binary. Compiler verbs and agent verbs share a namespace, so you stay in flow whether you're shipping a program or asking it questions.
- $ voss initScaffold a new project
- $ voss run app.vossCompile and execute
- $ voss compile app.vossEmit readable Python
- $ voss check app.vossType-check without running
- $ voss do "summarize this PR"One-shot agent task
- $ voss chatInteractive REPL
- $ voss doctorDiagnose your environment
Install.
Voss runs on Python 3.11+. Native binaries are on the roadmap.
Harness
A coding agent that runs on your subscription.
For AI-first developers shipping AI in their products. The Voss harness reuses your Claude Pro or ChatGPT login, runs sandboxed in your repo, and ships features without a second API bill.
Docs
The full reference is coming.
Language reference, runtime API, CLI deep-dive, and migration guides from raw Python. For now the PRD is the source of truth.