Software, by reflex. Say it, and the right small program runs when it is sure. A reflex is a recipe: written
once, shared, improved by everyone. The program and its arguments are chosen by Jev,
TypeSafe AI's System One classifier, which answers closed questions with calibrated probabilities. The questions,
and the gate that decides whether the call runs, confirms or asks, are evoke's.
$ evoke "kill the lights in the den"
lights room="den" state="off" 0.85
den lights off
curl -fsSL https://evoke.build/install.sh | sh # the CLI, on macOS and Linux
npm install @evoke-build/evoke # the SDKOne core, three ways in: a CLI you talk to, a package manager that installs reflexes from git, and a TypeScript SDK that puts the same decisions inside your app.
Saying what you want is easy. Trusting what runs is the hard part. An action taken on a guess costs more than one that never ran, so trust needs three things. Every value comes from you: from what you said, or from a list you own. The confidence is a number that means what it says, so you can set a bar on it. And what cannot be undone asks first, every time.
That is the idea behind a reflex. Your words pick a program. They never write one. Jev answers closed questions
about your sentence, every answer a calibrated probability, and the weakest one decides. evoke is the first
implementation of this idea: one core, a CLI, a package manager and a TypeScript SDK, with Jev as its first
engine and the design bound to none. The whole idea, with real sessions, at
evoke.build.
The value of all this is what it refuses to do. Three real sessions where nothing ran on a guess:
$ evoke "restart the computer"
power action="restart" · destructive · weakest: action 0.97
Really restart now? [y]es [n]o [t]each > n
[2]
$ evoke "set the volume to 150 percent"
How loud, in percent? 150 percent is outside 0–100 > 40
volume level="40" 0.93
volume set to 40%
$ evoke "make it cosy"
lights 0.45 · none 0.40 · timer 0.10 · volume 0.05 · route floor 0.50
[2]
Sure at 0.97, and it still asked: a destructive reflex always does. A value outside its range never reached a
program, and the range was on the line. A sentence that fits nothing ran nothing, and showed the ranking that
says why. evoke try puts every judgment on the table, and a bar is a line in a file you own.
Install, then the first ten minutes. Then write a reflex, or put decisions in your app. The whole manual is at evoke.build.
| Directory | Holds |
|---|---|
| crates/ | The Rust workspace: evoke-core, the rules with no engine and no I/O; evoke-adapters; evoke-wasm, the SDK's boundary; evoke, the CLI |
| sdk/ | @evoke-build/evoke: TypeScript over the same core, compiled to WebAssembly |
| spec/ | The executable spec — schemas, golden vectors, transcripts — and the tests of both hosts |
| reflexes/ | The first-party collection, published as evoke-build/reflexes |
| docs/manual/ | The manual, as published at evoke.build |
| site/ | evoke.build: the landing page, the install script, and the manual rendered by mdBook |
Tools and tasks are pinned by mise. Run mise install, then mise run lint and
mise run test. Those two are what CI runs; mise tasks lists the rest. bin/dx <command> runs a command with
the same toolchain inside an OrbStack machine named devbox, for those who develop in one. The tests need no key.
To decide against Jev, export TYPESAFE_API_KEY. To decide without one, the replay adapter answers from a file
you write: Testing.
Apache-2.0; the collection under reflexes/ is MIT.
Trademarks. Except for displaying the licence details and identifying us as the origin of the software, you have
no right under the licence to use our trademarks, trade names, service marks or product names: the name evoke, the
name evoke.build and the logo. A fork takes its own name and logo. TRADEMARK.md.