gawkbot

gawkbot is an open-source, self-hosted AI bot framework built in Go that runs entirely on your own machine using the AI coding runtime you already have — Claude Code, Codex, Opencode, or a local model. It was created by Najmuzzaman Mohammad and positions itself explicitly as an open-source alternative to xAI's Grok Bot, running on your keys instead of a vendor's cloud. The project reached the top of Hacker News Life of Product Week and has accumulated over 1,300 GitHub stars since its March 2026 creation.

What It Is

gawkbot is a multi-agent automation framework that turns natural-language workflow descriptions into always-on AI bots, each with its own working directory, tool allowlist, browser, and schedule. Rather than a single monolithic assistant, each bot is scoped to one workflow — expense filing, inbox triage, invoice reconciliation, file renaming — and when it finishes the job, it builds a microapp so you have a dashboard to review the outcome. The name is a deliberate parody: "gawk" means to stare blankly, which is what you do at the dashboard after the bots have already done the work.

How the Bot-and-Microapp Workflow Works

Every gawkbot ships with six components out of the box:

• The app — a real UI screen built live as the bot is created, reading and writing actual workspace data
• Routines — cron-style scheduled runs with versioned prompts and per-run transcripts
• Tools — self-authored capabilities the bot can teach itself from chat
• Knowledge — Wikipedia-style pages about the bot with citations back to source data
• Data + integrations — per-bot typed tables plus 1,200+ integrations via Composio
• Approval gate — reads are free; every write, send, commit, or delete waits for a human tap

Bots can consult each other: mention a second bot in a conversation and the first goes off to coordinate rather than pulling it into the thread. The build streams live in front of you — screen, routine, and tools assembling in real time.

Permissions and Trust Model

The permissions design is a first-class feature, not an afterthought. Each bot operates in its own directory with its own tool allowlist, so blast radius is per-bot rather than per-workspace. Approvals are per-action and do not persist — a blanket grant is not possible. Permissions that stop being used expire rather than accumulating silently. Every tool call is written to a local log. The project states that pointing gawkbot at a local model means nothing leaves the machine at all.

Setup Path and Runtime Options

Installation is a single npx gawkbot command; the browser opens, you verify your runtime, name your workspace, and describe your first workflow. A global npm install and a source build (requiring Go and Bun) are also supported. Runtime flags let you override the AI provider, change the web UI port (default 7891), or run multiple isolated workspaces side by side. Supported providers include Claude Code, Codex, Opencode, Ollama, Hermes Agent, and OpenClaw HTTP endpoints, all switchable via CLI flags or environment variables.

Update: v0.238.1

The latest release is v0.238.1, published 8 September 2026, with the repository last pushed on 9 September 2026. The project is under active development — the GitHub README includes a live claim-status table tracking shipped vs. planned features, and the comparison page against Grok Bot is dated 1 September 2026 with a standing invitation to file issues for any outdated rows. One planned item — packaging the routine runner into the npx/desktop installs — is listed as upcoming. The project is licensed under the Sustainable Use License, which permits free self-hosting for internal business and non-commercial use but prohibits reselling gawkbot as a hosted product.