dual-pane launchers · not a model runtime

grok-tua / tok-tua

Opens a coding CLI next to a live health and spend pane. You can see that Headroom → LiteLLM operates before you start a session.

View on GitHub v0.2.1 ai-gateway v0.2.1 · MIT · public

two binaries, one repo

They import each other. Split them and both break.

Left pane is the agent. Right pane is the door: Headroom, LiteLLM, spend, CLI versions. Coding takes ~80%. Meters take ~20%.

grok pane

grok-tua

Grok CLI on the left. Stack, quota, fleet CPU/RAM/GPU, and the active session’s burn on the right.

./bin/grok-tua → any CLI

tok-tua

Pi, OpenCode, Claude Code, Codex, Cursor — same dual-pane. Launch already aimed at Headroom.

./bin/tok-tua --cli pi → the door

ai-gateway

Headroom → LiteLLM glue these launchers assume. They do not route tokens. They show whether the bus is up.

the1truedan.github.io/ai-gateway → the house

fast-models

Shared weights and caches. A launcher without a pool is a dashboard with nowhere to work.

the1truedan.github.io/fast-models → thrift

Headroom

Default front door the pane pings first. Compresses context so a long agent session does not dump a novel into a paid model.

github.com/chopratejas/headroom → spend

LiteLLM

The metered source of truth in the pane: today= and window= dollars, plus which worker alias is live.

github.com/BerriAI/litellm →

what it looks like

Real sessions. One LAN address blacked out.

grok-tua dual-pane: Grok CLI left, stack and spend right — grok-tua — Grok CLI + fleet dashboard

tok-tua dual-pane: OpenCode left, gateway and credits right — tok-tua — OpenCode + gateway / credits

try it

Defaults to localhost. Nothing phones a stranger’s machine.

git clone https://github.com/the1truedan/grok-tua-tok-tua
cd grok-tua-tok-tua
./bin/tok-tua --cli pi --model manager-code
./bin/grok-tua

how it came to be

The door worked. Starting work was the bottleneck.

Once the local gateway was up, the daily question became: which CLI, which model tier, is Headroom alive, am I about to burn cloud credit by mistake?

grok-tua and tok-tua grew out of caregiver-hours shipping: cloud models helping sketch routing and smoke checks, the local repo holding the real scripts. They were extracted from the monorepo so the launchers could version on their own.

July 2026 — monorepo tooling (v0.2 era). Late July / early August 2026 — public extract. They live in one repo on purpose: each imports a bit of the other.

They are how humans and agents open the ai-gateway door. They are not the door.