GrokBot + Jev

Source

codila (@0xCodila) — x.com/0xCodila/status/2101433560796467348 (19 Sep 2026). ~1.3k likes / ~3.3k bookmarks / ~175k views at capture. ~16 s clip. Quotes his own Jev article ( x.com/i/article/2077744776565149697 , 18 Sep).

Open-source lab: github.com/Bodila51/grok-bot-jev (MIT). GitHub user is Bodila51; the tweet is @0xCodila. Same person unless proven otherwise.

Jev + GrokBot is the best AI agent system I’ve built in my life It just made my setup CHEAPER and FASTER than what 95% of people are running... setup takes literally 7 minutes

Loop he wants: prompt → Grok Bot → Jev decision → Grok Bot execution → result.

Skepticism (read this first)

This is the useful part. The tweet is a recirc. The README is the product. Don’t install on the back of the overlay.

• Jev is not a new Grok Bot. It cannot write, browse, or remember. An agent that only ran Jev would sit silent. You still need the desk you already have.
• There is no native switch. This is a pasted skill + a Python router on Agent Computer. If the bot ignores the skill, nothing is gated. The package cannot intercept a wake it wasn’t called on, and it cannot cut the cost of waking Grok Bot in the first place.
• “Best system / 95% / literally 7 minutes” is taste, not a bench. Clone + key + dry-run is a short sitting. Trusting active is a week of logs.
• CHEAPER is unproven on the meter that matters. Their own local A/B: Grok Bot weekly usage 37% → 38%, then 39%. They say that is not a token-savings claim. The 13× clip (53.8s → 4.1s) is one search with a top-five cap and two Jev calls (~$0.0004). Fewer pages was the design, not a miracle.
• Extra vendor, extra call, extra failure mode. TypeSafe key / waitlist. api.typesafe.ai down or wrong and you either stall or bypass. Jev can be wrong, unavailable, or slower than the work you skipped.
• It does not replace herdr, Chief, or Require Approval. Different meters, different jobs. ask_human is a classifier output, not Auto-review. Active mode still must not send / spend / publish.

Try it anyway? Yes — shadow, one research or browse bot you already run, key not in chat. If logs are mostly “allow / ignore,” don’t flip active. That’s the whole recommendation.

Should you try this (you already have Grok Bot)

Yes, in shadow mode, if the weekly meter is dying on browser tabs, “one more search,” same-approach retries, or spawning a second bot for a job the first one already did. That is the same furnace as /grokbotlimits.

Not instead of Chief / Require Approval / herdr. Jev is a cheap fork, not a new desk. The README is more honest than the tweet: one local A/B run; Grok Bot weekly meter even went 37% → 38%. They explicitly do not claim token savings.

Blockers before you start: a TypeSafe API key (typesafe.ai / docs), the bot’s Agent Computer able to pip-install, and a skill the bot will actually obey. There is no native Jev switch in Grok Bot.

What Jev is (so the 7 minutes make sense)

Jev is TypeSafe AI’s System One model (Kahneman: fast, not deliberative). It is not an LLM. It does not write a sentence. You send state + fixed questions; it returns typed answers with probabilities in ~70–500 ms. Three primitives:

Price TypeSafe published: on the order of $0.042 per million input tokens, output free because there is no generated text. That is why “ask Jev first” can be cheaper than another Grok tool loop — if the classification actually skips work.

An agent that only ran Jev would sit silent. Grok Bot still owns memory, computer, writing, routines. Jev owns the fork: reuse cache / stop retry / chat only / cap research / allow subagent / ask human.

His 7 steps (tweet), with the missing fences

1. Open TypeSafe, create an API key. Keep it off chat paste. Store it like /keychain (or Grok Bot’s secure field in step 2). Not in the repo, not in a screenshot.
2. Tell Grok Bot: store TYPESAFE_API_KEY in the secure field. Labels in chat, secret in the vault. The lab’s src/secrets.py reads only that env var.
3. Prompt Grok Bot: install typesafe-sdk on Agent Computer + smoke system_one (Choice). Better: clone the reference lab so you get the router, not a one-off script:

   git clone https://github.com/Bodila51/grok-bot-jev.git
   cd grok-bot-jev
   python3 -m venv .venv
   .venv/bin/pip install -r requirements.txt
   cp config.example.yaml config.yaml
   # TYPESAFE_API_KEY from the secure field / env — do not commit
   .venv/bin/python scripts/dry_run.py

   No-network smoke: set enabled: false in config.yaml and run dry-run anyway.
4. Build the usage lab (router, dry-run, config, logs) — or clone the GitHub above. Clone. Don’t vibe a second router.
5. Add skill jev-usage-router: before browser / research / retry / extra bot → call the router, honor route.action. Copy skill/jev-usage-router.SKILL.md into a Grok Bot skill. Pass task metadata and short context. Never credentials.
6. Stay shadow first. Decisions are logged and advisory. Read logs. Kill switch: bypass jev in the prompt, or enabled: false in config. The router cannot force a bot that ignores the skill to stop.
7. Flip active only when logs look sane. Grok Bot obeys the route. Jev decides. Grok Bot executes. Humans still own send / delete / spend / publish / account changes — same as Require Approval. The README says the router never does those without a human path.

“7 minutes” is the tweet. Clone + key + dry-run + skill paste is a short sitting. Trusting active is not.

What the lab actually returns

Explicit actions (from the README): reuse_cache, stop_retry, run_deterministic, chat_only, research_capped, allow_subagent, ask_human.

Conservative on purpose. Logs decisions, not the API key. Does not execute browser, research, payments, or messages. Does not change Grok Bot’s foundation model. Does not route Cursor models. No pre-wake hook — the bot still has to wake and call the router, so this cannot save the cost of the wake itself.

Overlay vs the repo (don’t mash)

How this maps here

Week-one sitting (if you try it)

1. TypeSafe key → secure field / Keychain. Never in chat.
2. Clone Bodila51/grok-bot-jev on the Agent Computer. Dry-run.
3. Paste the skill. Set mode: shadow.
4. One bot you already use for research or browsing. Not the Chief, not X-publish.
5. A week of logs. If it mostly says allow_subagent / ignore, don’t flip active.
6. Active only on that bot. Kill switch rehearsed once (bypass jev or enabled: false).

What we keep vs skip

• Keep: Jev as a fork before expensive tools; shadow → logs → active; labeled actions; kill switch; skill must honor the route; human path on irreversible work; clone the lab.
• Verify: TypeSafe key / waitlist; whether Agent Computer can reach api.typesafe.ai; whether the skill actually fires before browser.
• Skip: “best system / 95% / 7 minutes”; flipping active on day one; a Jev personality bot; pasting keys; treating the 13× clip as a product SLA; expecting the weekly Grok Bot meter to drop because of one A/B.

Related on this site

• GrokBot limits — the meter this is trying to spare
• Tips — Require Approval still wins over a classifier
• CEO · field manual
• Keychain — how to store TYPESAFE_API_KEY
• Domain harness — outer control system; this is one cheap knob
• Codex Fast vs Medium — another usage knob, different product

Primary: @0xCodila — Jev + GrokBot · grok-bot-jev · docs.typesafe.ai