gawkbot

open source grok bot. Every agent on your machine becomes a gawkbot your Chief of Staff manages, and they all ask you from the Mac notch or your iPhone.

https://github.com/user-attachments/assets/84c4b8cf-fb5d-4e9b-b720-213648c2e20b

gawkbot lets anyone turn their manual workflows into microapps across 1200+ integrations in minutes. Describe the job in one sentence — or demo it once on a call — and your AI builds the bot that runs it: its own screen, its own schedule, its own tools, with a human approval gate on everything it sends. Runs local, on your machine, on your account.

Already running Claude Code, Codex or Opencode? gawkbot finds them, adopts them, and hands them to your Chief of Staff. When any of them needs you, a little gawkbot pops out of your Mac's camera notch. Answer with one key, talk back with your voice, or reply from your iPhone.

gawkbots automate your menial work via AI models and build you microapps to manage the outcome, so that you have a false sense of control.

grok (verb) — to understand something profoundly and intuitively. gawk (verb) — to stare openly and stupidly.

It is named after the second one. Not after the bots. The bots are working. You are the one with the dashboard open.

The open source Grok Bot

gawkbot is an open source Grok Bot: always-on AI bots on your own machine instead of xAI's cloud, free, on the coding bot you already pay for, with an approval gate on every external action. The honest version, row by row:

Full comparison, where Grok Bot is better, and the other open source alternatives (Rakazo, OpenMausBot, OpenBot): gawk.bot/open-source-grok-bot. Website: gawk.bot.

Get Started

Prerequisites: one bot CLI, signed in — Claude Code by default, or Codex CLI / Opencode. The first-run screen verifies your runtime before anything else happens.

Download for Mac: gawk.bot/download.html resolves the latest signed .dmg; every build is also on the GitHub Releases page. Open it, drag gawkbot into Applications, and launch it. You verify your runtime, name your office, and hand off your first workflow — you land on your first bot being built, live.

On your iPhone: the companion app lives in apps/ios and you build it with Xcode. Pair it once by scanning the code in Access & Health. See On your iPhone below.

Building from source (requires Go and Bun):

git clone https://github.com/najmuzzaman-mohammad/gawkbot.git
cd gawkbot
cd web && bun install && bun run build && cd ..
go build -o gawkbot ./cmd/wuphf
./gawkbot

Routine execution runs on a small sidecar service (agent/). The broker finds and supervises it automatically on source checkouts (set WUPHF_AGENT_DIR to point elsewhere, or WUPHF_AGENT_URL if you manage it yourself).

What you get

Every bot ships with all six. Not a chatbot in a trench coat.

If your workflow names a system that is not connected yet ("audit our HubSpot"), gawkbot asks before building — build against live workspace data now, or hold while you connect. It never silently re-scopes your job.

Every agent you already run

gawkbot looks for the agents already on your machine. It checks your PATH, the agents' config folders and the process list, and it never launches anything to look. It knows Claude Code, Codex, Opencode, Gemini CLI, Cursor CLI, Aider, Goose, Amp, Qwen Code, Copilot CLI, Crush, Droid, Ollama, MLX-LM, Exo, OpenClaw and Hermes gateways, and the Cursor and Windsurf IDEs.

Adopt one from Settings → Agents on this machine, or ask the Chief of Staff to do it (it asks for your approval first). An adopted agent becomes a gawkbot on the team. It runs the agent's own CLI on your own sign-in, and gawkbot's keys are kept out of its environment.

Every bot is tagged with who made it (yours, hired by CoS, adopted, imported) and where it runs (this machine or elsewhere), so you can always tell your bots apart from the ones reporting in from somewhere else.

The notch

On a Mac, the Chief of Staff lives in the camera notch. You get no desktop widget and no dock badge.

• It moves to tell you what's going on. Each bot shows a mood: working, idle, needs you, error or done. Bots peek out now and then just for fun. When several need you, they pile up on the notch, and if you keep them waiting they start chatting with each other.

• It sounds like what happened. Every kind of event has its own cartoon cue: a slide whistle and a boing for a question, a bell for an approval, a sad trombone for an error, a xylophone ta-da for a finished job. The trackpad taps when you hover.

• You do it all from the keyboard. ⌃⌥Space opens it from anywhere. Then:

  Key	Does
  J / K	Move between questions
  1–9	Pick an answer
  ↵	Take the recommended answer
  R	Reply in your own words
  M	Message the Chief of Staff
  V (hold), ⌥V while typing	Talk instead of typing
  O	Open the full app; close its window to come back to the notch
  Esc	Close

• You can just talk. Hold V and say it. Speech is turned into text by macOS, and nothing is sent until you press send.

On your iPhone

apps/ios has the same inbox for your phone: every bot's question in one list, hold-to-talk replies, the same sounds and haptics, and hardware-keyboard shortcuts. It reaches bots running in the cloud as well as the ones on your Mac. Build steps are in apps/ios/README.md.

Setup prompt (for AI bots)

Paste this into Claude Code, Codex, or Cursor and let your bot drive the install:

Set up https://github.com/najmuzzaman-mohammad/gawkbot for me. Read `README.md`
first, then build from source (`go build -o gawkbot ./cmd/wuphf`) and run
`./gawkbot` — the web UI opens at http://localhost:7891.

Walk the onboarding: verify the runtime, name the office, and start the first
workflow. Confirm you land on a bot being built (a live build feed beside a
chat), and that when it finishes the bot shows tabs for UI, Routines, Tools,
Data, Knowledge, and Integrations.

For agent conventions read `AGENTS.md`; for internals read `ARCHITECTURE.md`;
for forking read `FORKING.md`.

Options

Local models and custom endpoints

For custom OpenAI-compatible endpoints (LiteLLM, local proxies, Ollama):

WUPHF_OLLAMA_BASE_URL="http://127.0.0.1:20128/v1" \
WUPHF_OLLAMA_MODEL="openai/gpt-5.4-mini" \
gawkbot --provider ollama --no-open

--provider opencode shells out to the opencode CLI binary; MLX-LM and Ollama can be set up from the first-run screen with no cloud key at all.

Other runtimes

Already running Hermes Bot or an OpenClaw gateway? Point bots at them with --provider hermes-agent (default http://127.0.0.1:8642/v1) or --provider openclaw-http (default http://127.0.0.1:18789/v1). Endpoints, models, and auth are overridable via WUPHF_HERMES_AGENT_* / WUPHF_OPENCLAW_HTTP_* env vars or provider_endpoints in config.

Memory: the company brain

gawkbot ships with built-in memory — no backend choice, no A