sense-lang 0.20.0

A programming language for probabilistic and autonomous computation — where a model's answer, a staged side effect, a permission check, and a human approval gate are things the language itself understands, not conventions a library hopes you follow.

Website · Documentation

Install

pip install sense-lang

Requires Python 3.10+. No other runtime dependency for the interpreter itself.

Optional extras, only needed if you use the matching feature:

pip install sense-lang[anthropic]   # inference("anthropic", ...)
pip install sense-lang[openai]      # inference("openai", ...)
pip install sense-lang[mcp]         # connect_mcp(...)

Quickstart

def fib(n) -> Int:
    if n < 2:
        return n
    return fib(n - 1) + fib(n - 2)

i = 0
while i < 10:
    print(fib(i))
    i = i + 1

sense run fib.sns
sense repl

Reasoning about something never talks to a vendor SDK directly — it asks whichever model is currently delegated to, and reads back a structured Answer, never a plain string:

model reasoning = inference("mock", "demo")   # offline, no API key needed
set delegation = reasoning

result = ask("what database should we use?")
print(result.value)
print(result.confidence)

A side effect that changes the outside world gets its own lifecycle — calling it never runs the effect; only .commit() does:

irreversible action send_mail(to: String, body: String) requires email.send:
    print("sending to " + to)

policy: allow email.send

mail = send_mail("alice@example.com", "hi")   # nothing sent yet
mail.verify()
mail.commit()                                  # only now

Full walkthroughs, the language reference, and the safety model live at docs.sensecode.dev.

What's in the box

• ask/Model/Answer — vendor-agnostic reasoning, with confidence and source structurally distinct from a plain value.
• action — prepare → verify → commit, enforced by the interpreter; a reversible action must declare how to undo itself.
• policy/requires — capability-gated permissions, scoped exactly like variables.
• agent — a real runtime entity with identity, state, and a genuine pause/resume lifecycle.
• tool/skill/connect_mcp — capability-checked functions a model can be handed to call on its own.
• memory/persistent memory — session-scoped and durable, versioned key/value storage.
• import python "module" — reach into any installed Python package directly; the interpreter is itself written in Python.

Development

Contributing or building from source:

git clone <repo-url>
cd sense
python -m venv .venv
.venv/Scripts/python -m pip install -e .[dev]   # Windows
# source .venv/bin/activate && pip install -e .[dev]   # macOS/Linux

pytest -q

License

MIT — see LICENSE.