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 agent 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.
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.
gawkbot is an open source Grok Bot: always-on AI bots on your own machine instead of xAI's cloud, free, on the coding agent you already pay for, with an approval gate on every external action. The honest version, row by row:
| Grok Bot (xAI) | gawkbot | |
|---|---|---|
| Price | Bundled with SuperGrok and Cursor paid plans | Free. No account, no seats, no usage fees |
| Source | Closed | Public, Sustainable Use License |
| Runs on | xAI's cloud | Your machine, with your keys |
| Models | Grok, chosen for you | Claude Code, Codex, Opencode, local models, Hermes, OpenClaw |
| Bot computers | One per bot, in the cloud | Not yet. Per-bot directory and tool allowlist on your machine |
| Approvals | Bots act on your accounts around the clock | Every send, commit, purchase, and delete waits for your click |
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.
Prerequisites: one agent CLI, signed in — Claude Code by default, or Codex CLI / Opencode. The first-run screen verifies your runtime before anything else happens.
npx gawkbotThat's it. The browser opens, you verify your runtime, name your office, and hand off your first workflow — you land on your first agent being built, live.
Prefer a global install?
npm install -g gawkbot && gawkbotBuilding 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
./gawkbotRoutine 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).
Every agent ships with all six. Not a chatbot in a trench coat.
| Part | What it is |
|---|---|
| The app | A real screen, built live in front of you. Reads and writes real workspace data. |
| Routines | "Every Monday 9:00." Versioned prompts, run history, a transcript per run. New agents get a starter weekly routine from the workflow you described. |
| Tools | Self-authored. "Score a lead." "Post to #ae-handoffs." Teach more in the agent's chat. |
| Knowledge | Wikipedia-style pages about the agent, every claim cited back to its source. |
| Data + integrations | Its own typed tables, plus 1200+ integrations. Connect once; every agent shares it. |
| Approval gate | Reads are free. Writes are held until you tap approve. Then it runs 24x7. |
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.
Paste this into Claude Code, Codex, or Cursor and let your agent drive the install:
Set up https://github.com/najmuzzaman-mohammad/gawkbot for me. Read `README.md`
first, then run `npx 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 an agent being built (a live build feed beside a
chat), and that when it finishes the agent 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`.
| Flag | What it does |
|---|---|
--provider <name> |
Runtime override (claude-code, codex, opencode, ollama, hermes-agent, openclaw-http) |
--no-open |
Don't auto-open the browser |
--web-port <n> |
Change the web UI port (default 7891) |
--workspace <name> |
Use a specific workspace for one command (does not change the active workspace) |
--unsafe |
Bypass agent permission checks (local dev only) |
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.
Already running Hermes Agent
or an OpenClaw gateway? Point agents 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.
gawkbot ships with built-in memory — no backend choice, no API key. Your
workspace state lives in local files you can cat: agent knowledge, run
transcripts, and the company brain under ~/.wuphf/. Knowledge pages are
synthesized with citations back to their sources, so you can check the
receipts on anything an agent claims.
gawkbot init # First-time setup
gawkbot share # Invite one team member over Tailscale/WireGuard
gawkbot shred # Delete workspace state and reopen onboarding
gawkbot workspace list # Run multiple isolated workspaces side by side
gawkbot workspace switch <name> # Flip the active workspaceTwo ways to invite a teammate, both from the CLI:
Private network — Tailscale or WireGuard. Both machines on the same mesh; the invite never leaves the network:
gawkbot sharePublic tunnel — no shared network needed. The broker can spin up a
Cloudflare quick tunnel (POST /api/share/tunnel/start; the trycloudflare
URL is paired with a 6-digit passcode, invites are one-use and expire in 24
hours, and the join handler is rate-limited per source IP). cloudflared
ships with the npm install (pinned SHA256 per platform). The one-click
button for this is being resurfaced in the operator shell — until then the
endpoint is the path.
For the full walkthrough, see Share gawkbot With a Team Member.
Agents act through two providers — pick whichever fits:
- One CLI (default, local-first): actions execute through a local CLI on your machine; credentials never leave it.
- Composio (cloud-hosted OAuth): connect Gmail, Slack, HubSpot, and the rest of the 1200+ catalog from any agent's Integrations tab. Connections are shared across the office.
Either way, the approval gate holds every external write until you approve it.
gawkbot can send anonymous product analytics and session recordings (with typed text masked) to help us improve it. This is optional, controlled by you, and off unless a PostHog key is configured — a stock source build and every fork ship with no key, so they never phone home.
Two independent toggles (onboarding and Settings), both on by default, both reversible at any time:
- Product analytics — anonymous usage events: which flows are used, where people get stuck, error counts. We send counts and shapes only, never your content (no message text, task titles, customer data, or secrets).
- Session recording — recordings mask everything you type
(
maskAllInputs: true): passwords, API keys, and any form field are obscured. We capture layout, clicks, and navigation to fix rough edges.
No autocapture, no cookies (localStorage only). Self-hosted operators can
point at their own PostHog (WUPHF_POSTHOG_KEY / WUPHF_POSTHOG_HOST) or
leave the key unset to keep gawkbot fully dormant. Full taxonomy and policy:
docs/specs/product-analytics.md.
| One agent per workflow | Small enough to read in a minute, real enough to do the whole job — instead of one giant assistant that does everything badly. |
| You watch it get built | The build streams live: the screen, the routine, the tools, assembling in front of you. |
| Honest by default | No connected data → the app says "simulated" in plain text. Missing integration → it asks before building. Every knowledge claim carries a citation. |
| Approval gate | No email, Slack post, or CRM write leaves without a human tap. |
| Local | Runs on your machine, on your keys. Workspace state is files you can cat. |
| Cost you can see | Settings shows exactly what your agents have spent — dollars, tokens, runs. A typical agent build lands in the $1–2 range on Claude Code. |
| Price | Free to self-host (Sustainable Use License, your API keys). |
Every claim in this README, grounded to the code that makes it true.
| Claim | Status | Where it lives |
|---|---|---|
| Describe a workflow → agent builds live with a streaming activity feed | ✅ shipped | web/src/operator/surfaces/AppBuilderChat.tsx, web/src/components/apps/AppActivity.tsx |
| Onboarding hands the first workflow straight into the build | ✅ shipped | web/src/operator/firstWorkflowSeed.ts, web/src/operator/OperatorApp.tsx |
| New agents get a starter weekly routine from the described workflow | ✅ shipped | web/src/operator/surfaces/AppBuilderChat.tsx |
| Routines: broker-owned cron, versioned prompts, per-run transcripts | ✅ shipped | internal/team/scheduler_operator_routines.go, web/src/operator/routines/RoutinesTab.tsx |
| The broker spawns and supervises the routine runner | ✅ shipped | internal/team/agent_service_supervisor.go |
| Ask-before-building when a referenced integration is not connected | ✅ shipped | web/src/operator/builder/describedIntegrations.ts |
| Approval gate on external writes | ✅ shipped | web/src/operator/components/ApprovalPrompt.tsx, internal/team/broker_action_grants.go |
| Knowledge pages with inline citations | ✅ shipped | web/src/operator/surfaces/KnowledgeSurface.tsx |
| Usage readout — cost, tokens, runs | ✅ shipped | web/src/operator/surfaces/SettingsSurface.tsx (GET /usage) |
| Operator is the only front door; legacy office routes redirect | ✅ shipped | web/src/routes/RootRoute.tsx, web/e2e/tests/route-matrix.spec.ts |
| Per-agent typed tables (Data tab) | ✅ shipped | web/src/operator/surfaces/AppDataTab.tsx |
| Mix runtimes: Claude Code, Codex, Opencode, local models, Hermes, OpenClaw | ✅ shipped | internal/provider/, first-run pre-pick screen |
| Multi-workspace, share, and public-tunnel invites | ✅ shipped | cmd/wuphf/workspace.go, cmd/wuphf/tunnel.go, internal/workspaces/ |
| Pixel-art agent portraits (every agent has a face) | ✅ shipped | web/src/components/ui/PixelAvatar.tsx |
| Routine runner packaged into the npx/desktop installs | 🔜 planned | supervisor adopts it the moment it ships |
Legend: ✅ shipped · 🟡 partial · 🔜 planned. If a claim and a status disagree, the code wins — file an issue.
Before you fork, run this prompt against the codebase with any AI coding assistant:
You are a cynical senior engineer evaluating whether to fork this repo as the
base for an agents-that-run-workflows product. No prior context — explore it
as you naturally would. Tell me: should I fork this, and what's your honest
take? Be specific: file paths, line numbers, actual evidence. "The docs are
bad" is useless. Under 500 words.
We run this ourselves before every release. If the AI finds something we missed, file an issue.
Grok is named for a verb that means to understand something completely and intuitively. That is a confident name. It promises you a machine that gets it.
gawkbot is named for a verb that means to stare at something openly, at length, and without taking in any of it. That is the honest name for what actually happens once the agents are running. The work is done. You are looking at a dashboard about it.
We are a parody of a bot company. We also ship, which is the part they did not see coming.