Consultant stack

Tool notes in Yonatan’s own English; the rest translated from Hebrew

Yonatan Gross

Tech & AI stack: Mostly tools for building software, AI and automations

AI systems engineer · Yonyon AI

  • Builds AI agents, automations and integrations that run in production for businesses
  • Tel Aviv
  • Independent since 2025

At a glance

  • 69 tools
  • 6 categories
  • 69 uses personally

Core stack

Picked only from the consultant's own words: daily use, years of use, "almost exclusively".

  • Claude CodeDaily use

    “The tool that built this site and most of my business.”

  • GitHub24 repos

    “Code, issues, CI and releases, in one place.”

  • Wispr FlowEvery day

    “Most of what I send an agent starts as speech, not typing.”

  • OrchestKitHis own project

    “The plugin that turns Claude Code into a team with roles.”

  • Anthropic API

    “Claude runs the agents, the site chat, and the ops bots.”

  • Docker

    “Local dev matches production because both are the same compose stack.”

  • Slack19 channels

    “Every bot, deploy and alert reports into a channel, not into a log file.”

“I build AI that runs itself, then keep it running.” The list here is the stack yonyon.ai runs on, as Yonatan published it on his stack page: “in three rings: my daily loop at the center, the infrastructure and models it runs on next, and the channels and dashboards that carry it outward.” All 69 tools, each with its dossier, are on yonyon.ai.

Details for every tool

In the consultant's own words, with a link to the source of each quote.

Copies the scored tools into your private My Stack. Nothing is shared.

Daily loop7
  • Claude CodeNot in Curalo yetUses it

    Why:Builds everything: “The tool that built this site and most of my business.” In practice: This site: 500+ pull requests written with Claude Code. The site chat runs on Claude with my knowledge base. My ops bots investigate deploys and post summaries to Slack.

    From the article ↗

  • GitHubNot in Curalo yetUses it

    Why:Where the work lives: “Code, issues, CI and releases, in one place.” In practice: 24 active repos across the org. Every pull request runs the full check suite before merge. Promotes to production are merged by an App, never by hand.

    From the article ↗

  • Wispr FlowNot in Curalo yetUses it

    Why:I talk, it types: “Most of what I send an agent starts as speech, not typing.” In practice: Briefs to my agents get dictated, not typed. Mixed Hebrew and English in one sentence still lands clean.

    From the article ↗

  • DockerNot in Curalo yetUses it

    Why:Everything runs in a container: “Local dev matches production because both are the same compose stack.” In practice: Local dev runs the same compose stack as production. Production deploys blue-green. The test stack is ephemeral, so integration tests never touch the dev database.

    From the article ↗

  • Google WorkspaceNot in Curalo yetUses it

    Why:Mail, calendar and docs: “Where the business mail and calendar live, readable by my agents.” In practice: My agents search my inboxes instead of me scrolling. Meeting slots are proposed from the real calendar, not guessed.

    From the article ↗

  • Codex CLINot in Curalo yetUses it

    Why:A second coding agent: “OpenAI's coding agent, for tasks where a second model is the better fit.” In practice: Some coding tasks go to Codex instead of Claude Code.

    From the article ↗

  • CandleKeepNot in Curalo yetUses it

    Why:A library my agents can read: “Books and my own notes, searchable by agents before they reinvent an answer.” In practice: My agents check the library before re-deriving a lesson.

    From the article ↗

Build and code20
  • OrchestKitNot in Curalo yetUses it

    Why:My agent operating system: “The plugin that turns Claude Code into a team with roles.” In practice: 106 skills and 37 agents, installed in every repo I work in. Its hooks refuse commits that repeat a mistake I already paid for.

    From the article ↗

  • StorybookNot in Curalo yetUses it

    Why:Every component, in isolation: “Where a component is proven before it reaches a page.” In practice: My agents query the catalog before writing a new component, so they reuse instead of duplicate. Components get verified without logging into the app.

    From the article ↗

  • shadcn/uiNot in Curalo yetUses it

    Why:The UI primitives: “Components I own the code for, not a library I am locked into.” In practice: Every primitive on this site started as a shadcn component I then made mine. My design system layers tokens and patterns on top of it.

    From the article ↗

  • CursorNot in Curalo yetUses it

    Why:A second pair of eyes: “A separate review lane, so one model does not grade its own homework.” In practice: Kept as a reviewer lane beside my main agent, on purpose. It only earns its seat if it catches a class of bug the other lane misses.

    From the article ↗

  • Next.jsNot in Curalo yetUses it

    Why:The web framework: “The framework this site and the client sites I build run on.” In practice: This site runs on Next.js 16, in English and Hebrew. The client sites I build run on it too.

    From the article ↗

  • TypeScriptNot in Curalo yetUses it

    Why:Types everywhere: “Every JavaScript repo I own is TypeScript, so mistakes fail the build, not the visitor.” In practice: Every JavaScript repo I own is TypeScript. A type error blocks the merge on this site.

    From the article ↗

  • ReactNot in Curalo yetUses it

    Why:Every interface I ship: “The UI layer under this site, my dashboard and the client sites.” In practice: This site, my dashboard and the client sites are all React 19.

    From the article ↗

  • Tailwind CSSNot in Curalo yetUses it

    Why:The styling layer: “Utility classes on top of design tokens, so light, dark and Hebrew all come from one source.” In practice: Every style on this site is a Tailwind utility or a token. Hebrew layout comes from logical properties, not a second stylesheet.

    From the article ↗

  • FastAPINot in Curalo yetUses it

    Why:The API behind my agents: “The Python API my platform and its agents run through.” In practice: My platform's API is a FastAPI app. Request and response models are typed end to end.

    From the article ↗

  • PlaywrightNot in Curalo yetUses it

    Why:End-to-end tests: “Real browsers click through the site before a change can merge.” In practice: The end-to-end suite runs on every pull request to this site. The hero has snapshot baselines in Chromium, Firefox and WebKit.

    From the article ↗

  • MotionNot in Curalo yetUses it

    Why:Animation: “The motion on this site, from section reveals to this orbit's dossier.” In practice: Section reveals and the dossier opening on this site are Motion.

    From the article ↗

  • RemotionNot in Curalo yetUses it

    Why:Video, written as code: “Videos composed in React, so a change is a commit, not a re-edit.” In practice: My creative toolkit renders video through Remotion.

    From the article ↗

  • VitestNot in Curalo yetUses it

    Why:Unit tests: “The fast test layer that runs on every push.” In practice: This site runs about 2,600 unit tests before a push goes out.

    From the article ↗

  • pnpmNot in Curalo yetUses it

    Why:Package manager: “Installs for every JavaScript repo, with a delay on brand-new releases.” In practice: Every JavaScript repo I own installs with pnpm. A package has to be a day old before it can land.

    From the article ↗

  • CodeRabbitNot in Curalo yetUses it

    Why:AI code review: “A reviewer that reads every pull request before I do.” In practice: Reviews pull requests in three of my repos.

    From the article ↗

  • RenovateNot in Curalo yetUses it

    Why:Dependencies, kept current: “A bot that opens the update pull requests so I only review them.” In practice: Keeps dependencies current in 12 of my repos.

    From the article ↗

  • uv + RuffNot in Curalo yetUses it

    Why:Python, fast: “Astral's installer and linter for every Python project I run.” In practice: Python installs and lint in my CI run through uv and Ruff.

    From the article ↗

  • TurborepoNot in Curalo yetUses it

    Why:The monorepo build: “Builds and tests only what changed, and remembers the rest.” In practice: My platform monorepo builds and tests through Turborepo.

    From the article ↗

  • Context7Not in Curalo yetUses it

    Why:Current docs for agents: “Agents read the library docs as they are today, not as they were at training time.” In practice: My agents check the current docs before writing against a library.

    From the article ↗

  • 21st.devNot in Curalo yetUses it

    Why:Component search: “A registry of ready components my agents search before building one.” In practice: My agents search it before writing a new component from scratch.

    From the article ↗

Infrastructure5
  • VercelNot in Curalo yetUses it

    Why:Ships the front ends: “Every client site and dashboard deploys here on push.” In practice: This site deploys on every merge, with a preview URL per pull request. Client sites run here too, each on its own project.

    From the article ↗

  • CloudflareNot in Curalo yetUses it

    Why:The edge in front of everything: “DNS, the firewall and email security in front of every site I run.” In practice: Every domain I run is managed and protected here.

    From the article ↗

  • TerraformNot in Curalo yetUses it

    Why:The servers are a file: “No server exists because someone clicked a button.” In practice: The CI fleet's size and shape live in version control, not in a console. A box can be rebuilt from the repo rather than from memory.

    From the article ↗

  • Cloudflare R2Not in Curalo yetUses it

    Why:Where the files live: “Object storage with no egress fee, which is the whole reason.” In practice: Client deliverables and generated media are stored and served from here. Public and private buckets are split, so nothing leaks by default.

    From the article ↗

  • UpstashNot in Curalo yetUses it

    Why:Rate limits without a server: “The serverless Redis that rate-limits this site's chat and API routes.” In practice: Rate-limits the chat and API routes on this site.

    From the article ↗

Models and AI21
  • Claude Agent SDKNot in Curalo yetUses it

    Why:Agents inside my API: “The same agent loop as Claude Code, running inside my own services.” In practice: My platform's API runs Claude agents through the SDK.

    From the article ↗

  • Vercel AI SDKNot in Curalo yetUses it

    Why:AI in the web apps: “How model output streams into the interfaces I build.” In practice: This site and my dashboard stream model output through it.

    From the article ↗

  • Anthropic APINot in Curalo yetUses it

    Why:The model behind the work: “Claude runs the agents, the site chat, and the ops bots.” In practice: The chat on this site answers from my own knowledge base. My ops bots read deploys and write the summary themselves.

    From the article ↗

  • OpenAINot in Curalo yetUses it

    Why:Second opinion and images: “Where I go when a different model is the better tool.” In practice: Marketing images and product cards are generated here. Kept in the fallback chain so one provider outage is not an outage.

    From the article ↗

  • Google GeminiNot in Curalo yetUses it

    Why:Video, images and voice: “Google's model family for everything that is not text.” In practice: Hebrew voice output uses Google's speech, pinned to a specific voice. Video and image generation sit behind a circuit breaker, so a failing provider is skipped instead of retried.

    From the article ↗

  • GroqNot in Curalo yetUses it

    Why:Transcription, fast: “Whisper at a speed that makes voice notes practical.” In practice: Short voice messages get transcribed here before anything reads them.

    From the article ↗

  • OpenRouterNot in Curalo yetUses it

    Why:One key, many models: “Lets me swap a model without swapping an integration.” In practice: Sits in my eval path: the cross-model matrix scores the same prompt on several vendors' models through one key.

    From the article ↗

  • HiggsfieldNot in Curalo yetUses it

    Why:Brand visuals and video: “Where the cinematic brand work gets made.” In practice: Brand images and reels for this business are produced here. Stills get approved before any video renders, so a bad frame never becomes a bad clip.

    From the article ↗

  • fal.aiNot in Curalo yetUses it

    Why:Image models on tap: “The generation backend behind the creative pipeline.” In practice: Sits behind a circuit breaker with a provider fallback chain, so one bad provider does not stall a pipeline.

    From the article ↗

  • DeepgramNot in Curalo yetUses it

    Why:Who said what: “Speaker separation on recorded calls.” In practice: Only runs on audio long enough to be worth it, so short clips cost nothing.

    From the article ↗

  • SonioxNot in Curalo yetUses it

    Why:Hebrew that survives: “The one transcriber that handles Hebrew and English in the same sentence.” In practice: Meeting audio tries Soniox first, because code-switched Hebrew breaks the others.

    From the article ↗

  • CartesiaNot in Curalo yetUses it

    Why:The English voice: “Low-latency speech for anything that talks back.” In practice: Handles English speech output, and takes over when a sentence mixes languages.

    From the article ↗

  • ElevenLabsNot in Curalo yetUses it

    Why:Music beds for the studio: “Wired into my creative studio for generated music; speech runs elsewhere.” In practice: Wired for music beds in my creative studio; the everyday speech chain runs on Cartesia and Google instead.

    From the article ↗

  • TavilyNot in Curalo yetUses it

    Why:Search built for agents: “Web results an agent can actually use, without the page noise.” In practice: My research skills call it instead of scraping pages into context.

    From the article ↗

  • NotebookLMNot in Curalo yetUses it

    Why:Turns docs into a briefing: “Where a pile of sources becomes something I can listen to.” In practice: Each release gets a notebook and an audio summary I listen to instead of reading.

    From the article ↗

  • Voyage AINot in Curalo yetUses it

    Why:Makes search understand meaning: “The embeddings behind my knowledge base.” In practice: The knowledge base finds answers by meaning, which is why the site chat can answer in my words.

    From the article ↗

  • LangGraphNot in Curalo yetUses it

    Why:Agent workflows: “Multi-step agent flows that can pause, resume and be inspected.” In practice: Multi-step agent flows in my platform's API are LangGraph graphs.

    From the article ↗

  • PipecatNot in Curalo yetUses it

    Why:Voice agents: “Chains speech-to-text, a model and a voice into one real-time conversation.” In practice: My voice agents are Pipecat pipelines in my platform's API.

    From the article ↗

  • FastMCPNot in Curalo yetUses it

    Why:My MCP servers: “The framework behind the MCP servers that give my agents typed tools.” In practice: Eight MCP servers I built give my agents their tools.

    From the article ↗

  • GrokNot in Curalo yetUses it

    Why:Another model in the mix: “xAI's model, one of the models I test side by side before picking one for a job.” In practice: I run Grok against the other models on the same tasks.

    From the article ↗

  • WhisperNot in Curalo yetUses it

    Why:Transcription on my laptop: “Speech to text that runs locally, so call audio never leaves the machine.” In practice: Client call recordings are transcribed on my own laptop.

    From the article ↗

Channels and distribution12
  • SlackNot in Curalo yetUses it

    Why:Where I read the business: “Every bot, deploy and alert reports into a channel, not into a log file.” In practice: Each channel carries one class of message, so alerts never drown the ops feed. Deploys, cost anomalies and client health post themselves here.

    From the article ↗

  • WhatsAppNot in Curalo yetUses it

    Why:The channel clients actually answer: “In Israel, email is where a thread goes to die.” In practice: Client voice notes get transcribed and filed against that client automatically.

    From the article ↗

  • ResendNot in Curalo yetUses it

    Why:Email that arrives: “Transactional mail without a deliverability project.” In practice: Order confirmations, shipping and invoices all send from here. Bounces and opens come back as webhooks, so a silent failure is visible.

    From the article ↗

  • Why:Booking without the back and forth: “A link instead of six messages about Tuesday.” In practice: Prospects pick a slot from a link, so I never negotiate a time by message. Each booking posts a webhook to this site, which sends the follow-up email.

    From the article ↗

  • XNot in Curalo yetUses it

    Why:Where I publish in English: “Short, technical, and usually about something that just broke.” In practice: Every post is drafted by an agent and approved by me before it goes out. The bot posts from its own account, never my personal one.

    From the article ↗

  • LinkedInNot in Curalo yetUses it

    Why:Where I publish long form: “Longer posts, for the people who actually hire me.” In practice: Long-form posts go here, drafted by an agent and edited by me. Nothing posts without me approving it first.

    From the article ↗

  • InstagramNot in Curalo yetUses it

    Why:Where the visual work goes: “Brand and client visuals, cross-posted from one place.” In practice: Visual work publishes from the same pipeline as everything else, not by hand.

    From the article ↗

  • ThreadsNot in Curalo yetUses it

    Why:A second English feed: “Same drafts, one more audience, no extra work.” In practice: Posts arrive here from the same approved draft, so it costs nothing extra to run.

    From the article ↗

  • TelegramNot in Curalo yetUses it

    Why:The bot channel: “Where machine chatter goes so it does not clutter a human channel.” In practice: Machine-to-machine notifications live here, away from anything a client can see.

    From the article ↗

  • DiscordNot in Curalo yetUses it

    Why:Where the communities are: “Less a tool I run, more where the people I learn from sit.” In practice: The account runs on the free tier: a place to learn, not something I run on.

    From the article ↗

  • Why:The long version: “Where a thing that needs 2,000 words goes.” In practice: One of the distribution targets in my content pipeline; the subscriber list itself lives in my own database.

    From the article ↗

  • RiversideNot in Curalo yetUses it

    Why:Recording and editing: “Where I record video, and where my agents help cut it.” In practice: My agents can search a recording's transcript and cut the edit.

    From the article ↗

Measurement and operations4
  • LangfuseNot in Curalo yetUses it

    Why:Every model call, traced: “Every prompt and output is traced, so I can show what an agent did and what it cost.” In practice: Every agent run is traceable end to end, with what it cost.

    From the article ↗

  • PostHogNot in Curalo yetUses it

    Why:What people actually do: “Product analytics without shipping my users to an ad network.” In practice: Loaded only after consent, and lazily, so it never costs a visitor page speed.

    From the article ↗

  • ProdigiNot in Curalo yetUses it

    Why:Prints and ships for me: “Print on demand, so I hold no stock.” In practice: An order routes straight to fulfillment, and the customer gets tracking without me in the loop.

    From the article ↗

  • SanityNot in Curalo yetUses it

    Why:Content for eight sites: “One CMS, eight projects, edited without a deploy.” In practice: Eight separate projects, one per site, all reachable from one place. This very section is edited here; changing a tool takes no deploy.

    From the article ↗

Q&A

What do you do, and for which clients?
“Beyond courses and workshops, I build and run AI in production: my own business's operating system runs on AI agents, with monitoring, alerts and incident handling.”
What do you trust AI with in client work, and what never?
“One engineer sets direction and reviews every change. The AI team does the repetitive work.”

Disclosure

OrchestKit is Yonatan's own open-source project. And on his stack page: “Some links are referral links.” (The links on this page go directly to the tools' websites.)

Some tool links on Curalo are affiliate links. That never changes a tool's score or rank, and consultants are not paid for their profiles.

Last confirmed October 25, 2026