Quickstart
decide asks a decision model questions about your data. A decision model answers with yes or no, a choice, or a score, each with its probability, and writes no text.
Install decide
Section titled “Install decide”brew install deepnoodle-ai/tap/decidego install github.com/deepnoodle-ai/decide/cmd/decide@latestDownload a binary for Linux, macOS, or Windows from the
releases, and put it
on your PATH.
decide --version prints the version you have.
Get a key
Section titled “Get a key”Pick one provider and export its key:
Jev by TypeSafe is the default. Get a key at typesafe.ai.
export TYPESAFE_API_KEY=...Clef runs on Cloudflare Workers AI. Create an API token that can use Workers AI in the Cloudflare dashboard.
export CLOUDFLARE_AUTH_TOKEN=...export CLOUDFLARE_ACCOUNT_ID=...export DECIDE_PROVIDER=cloudflareGPT-6 Luna runs on OpenAI’s Decisions API, in beta. Create a key in the OpenAI platform.
export OPENAI_API_KEY=...export DECIDE_PROVIDER=openaiGemini Decision runs on Google’s Gemini Decisions API, in beta. Create a key in Google AI Studio.
export GEMINI_API_KEY=...export DECIDE_PROVIDER=geminidecide models checks that the key works. decide sends what it asks
about, such as your diff or a file you name, to the provider you chose.
Ask about one line
Section titled “Ask about one line”Pipe in a line of text and run the built-in sentiment template:
echo "The new release fixed everything I cared about" | decide run sentiment
Output as text
$ echo "The new release fixed everything I cared about" | decide run sentiment Running sentiment on 1 line · typesafe jev-latest stdin:1 The new release fixed everything I cared about sentiment positive 100% ✓ 1 answered nothing flagged 200ms See these results again with: decide runs view 20261112-170305-d799
The answer is positive, with its probability. Your percentage may differ a little.
Ask about your own changes
Section titled “Ask about your own changes”In a git repository with unstaged changes, judge each changed file:
git diff | decide run code-risk
Output as text
$ git diff | decide run code-risk Running code-risk on 4 files · typesafe jev-latest AGENTS.md risk no 98% maintainability ━━━━━━━━━━━━ 4.0 of 4 Exceptionally clear and focused billing/invoice.go risk no 98% maintainability ━━━━━━━━━━━━ 4.0 of 4 Exceptionally clear and focused billing/refund.go ! risk yes 95% maintainability ━━━━━─────── 1.6 of 4 Understandable with some friction docs/integrations.md risk unsure 54% yes maintainability ━━━━━━━───── 2.3 of 4 Understandable with some friction ✓ 4 answered ! 1 flagged 100ms Flagged: billing/refund.go See these results again with: decide runs view 20261112-170246-b54a
billing/refund.go is flagged as a risk. This
recording is from the demo’s shop repository.
code-risk asks whether each file could cause security or data
problems, and how maintainable it is. A flagged answer is red and marked
with !.
To make a script or CI act on the answers, add --fail-on flagged. decide
then exits with code 2 when anything is flagged:
git diff | decide run code-risk --fail-on flaggedAsk about a screenshot, recording, or PDF
Section titled “Ask about a screenshot, recording, or PDF”Name an image, audio, video, or PDF file, and any template asks its questions about it. Each file is one item. Jev reads text only, so choose a provider that reads the file:
decide run code-risk diff.png --provider openaidecide run sentiment call.mp3 --provider geminiClef, GPT-6 Luna, and Gemini Decision read images. Only Gemini Decision reads audio, video, and PDFs. If your provider can’t read a file, decide stops before it sends anything and names the providers that can. See Images, audio, video, and PDFs.
What next
Section titled “What next”decide templates lists the twelve built-in templates.
decide is Apache 2.0. Made by Deep Noodle, after some very hard thinking.