CraftBot: A Self-Hosted AI Agent That Builds Its Own Tools
CraftBot is a self-hosted AI agent that builds custom SaaS tools while using them to automate tasks. Open-source, multi-LLM support, runs locally. Review +

The AI agent space is split roughly into two camps right now. On one side you have the hosted services - ChatGPT Pro, Claude with computer use, Manus, the wave of "do everything for you" cloud agents. On the other, a growing collection of open-source frameworks aimed at self-hosters who want the agent to actually live on their own hardware, talk to their own data, and not phone home every five seconds.
CraftBot sits firmly in the second camp - and it earns the distinction by trying to do something more ambitious than most self-hosted agents bother with: it doesn't just chat with you, it builds you tools, then operates them on your behalf.
What CraftBot actually does
The core idea: instead of generating answers, CraftBot generates and runs lightweight SaaS-style tools. Ask it to track your habits and it'll spin up a kanban board you can interact with directly, then keep operating that board over time as your routines change. Ask it to manage a PM workflow and it'll build a dashboard, wire in the integrations you authorise, and act on incoming events.
Under the hood, it works as a general-purpose agent harness with four things going on at once:
- Tool generation: CraftBot writes the React + REST scaffold for each tool you ask it to build. The frontend it calls a "Living UI system".
- Memory + learning: a RAG-based memory layer consolidates daily events and surfaces user preferences over time, so the agent's suggestions get more contextually accurate the longer you use it.
- Multi-step workflows: it can chain actions across tools, not just within a single one.
- Proactive suggestions: it watches what you do and offers next-step actions rather than waiting to be prompted.
Setup: about as painless as it gets
One of the things that drew me to CraftBot rather than rolling my own AutoGen or LangGraph harness was the install story. You need:
- Python 3.10 or later (most modern Linux + macOS setups already have this)
- Node.js 18+ only if you want the browser UI; the CLI mode works without Node
Then a single command does the rest:
python craftbot.py installThat installer handles the auto-start + background-service setup on Windows, macOS and Linux. On Ubuntu I had it running as a systemd unit in under five minutes. The conda / pip-only install paths skip the service automation - you run it manually via run.py if you'd rather manage your own process lifecycle.
Which models can you point it at?
For a self-hosted tool, the LLM compatibility list is what matters most. CraftBot supports:
- OpenAI (GPT-5 family + earlier)
- Anthropic Claude (Sonnet 4.6 + Opus 4.7 at time of writing)
- Google Gemini
- BytePlus
- Local Ollama models - this is the one that matters if your self-hosting goal is keeping data off third-party clouds
The local-Ollama path is genuinely first-class, not an afterthought. If you've got a Linux box with a recent NVIDIA GPU (or a decent Apple Silicon Mac), you can run CraftBot end-to-end without any of the agent's reasoning leaving your network. That changes the privacy calculus quite a lot - particularly if you're connecting it to email, Slack or Notion via the built-in OAuth integrations.
What's it actually useful for?
Three use cases stood out for me after a couple of weeks of running it:
- PM-style dashboards: the headline example. CraftBot builds you a kanban + a backlog view + a sprint timeline, then operates them. You can talk to it in natural language to triage tickets, move items, set priorities. It handles the rote bookkeeping while you stay in decision-making mode.
- Knowledge consolidation via RAG memory: drop documents, transcripts and notes into its data layer and it builds a queryable memory that gets sharper over time. Combine that with Ollama and you've got a private "second brain" that doesn't depend on any SaaS vendor.
- Workspace automation: the OAuth integrations cover Google Workspace, Slack, Notion, Zoom, LinkedIn, Discord and Telegram. Useful for stitching together small repetitive workflows that aren't worth writing custom Zapier flows for.
Caveats and limitations
A few things worth knowing before you commit a weekend to setup:
- GUI mode is deprecated: earlier versions supported desktop automation; the maintainers have explicitly dropped that. If you needed CraftBot to click around your screen, it won't do that any more.
- WhatsApp Web integration needs Playwright Chromium: the install is optional and fails gracefully if you skip it, but you need it specifically if WhatsApp is part of your workflow.
- Active development: at 354 GitHub stars and a 1.3.3 release in June, this is still very much an early-stage project. Expect API surface changes between minor versions; pin your installation if you're using it for anything mission-critical.
Should you self-host an agent?
If you're already running a home server, NAS or beefy Linux box and you've been eyeing the AI-agent space without wanting to feed every interaction to OpenAI's training corpus, CraftBot is one of the cleanest on-ramps I've found. The install story is painless, the local-Ollama support is real, and the "build me a tool" workflow is genuinely different from "chat with me about a topic."
It's not a polished commercial product. You'll bump into rough edges. But the trajectory is right, and the project is moving fast - the v1.3.3 release shipped on the day I'm writing this. Worth a weekend of your time if you've got the hardware sitting idle.