Local AI agents are having a moment right now.
Developers are increasingly uncomfortable with cloud-only assistants privacy concerns data leaving your machine, API costs that compound The push toward running AI locally is real and the tooling is catching up fast.
I spent time testing two tools that take very different approaches to this problem:
OpenClaw - a browser extension for local AI tasks lightweight and ready in 60 seconds
CraftBot - a full desktop AI agent built by CraftOS open source and capable of things a browser extension simply can't do
This isn't a sponsored comparison I tested both documented everything and I'll tell you exactly what I found.
What Is OpenClaw?
OpenClaw is a browser extension available in the Chrome Web Store and installable in about a minute It's built around their Cooper agent and comes with a pre-built Skills Library for common tasks.
The core promise: quick lightweight browser-native AI assistance If you want something running in 60 seconds without touching a terminal this is it.
Best for: Developers who want a fast low-friction AI tool that lives in the browser.
What Is CraftBot?
CraftBot is a different beast entirely.
It's a full desktop AI agent - Python-based fully open source available on GitHub (CraftOS-dev/CraftBot 276+ stars at time of writing) It runs locally on your machine connects to external services via MCP (Model Context Protocol) and has a feature called Living UI that lets the agent build its own custom tools on the fly.
The setup takes longer. But the capability ceiling is in a completely different league.
Best for: Developers who need local model support file management observability and serious agentic workflows.
Installation & Setup
OpenClaw
1-click install from the Chrome Web Store Ready in approximately 60 seconds No terminal required.
CraftBot
GitHub: github.com/CraftOS-dev/CraftBot — clone this to get started
CraftBot requires Python 3.10+ git and Node.js Once you have those:
git clone https://github.com/CraftOS-dev/CraftBot.git
cd CraftBot
python craftbot.py install
CraftBot handles everything automatically from there.
CraftBot installation — Step 1 of 3, installing core dependencies at 94%. The retro ASCII art is a nice touch.
The installer downloads all required packages. After that, CraftBot launches in TUI mode and begins downloading the embedding model:
Dependencies installed — CraftBot launches, downloads the all-MiniLM-L6-v2 embedding model, and signals [OK] Ready
Note: CraftBot recommends Python 3.10, but the installation handles dependencies regardless of your Python version.
On first run, a 6-step onboarding wizard walks you through configuration. Step 1 asks you to choose your LLM provider:
9 provider options including OpenAI, Google Gemini, Anthropic, DeepSeek, Moonshot, and Ollama (Local). I chose Google Gemini for initial testing.
Step 5 lets you select MCP servers — this is where CraftBot's depth shows:
MCP server selection — Gmail, Slack, GitHub, Notion, Google Calendar, Todoist, Obsidian, Brave Search, Playwright, and Filesystem. Each one you enable adds tools your agent can call.
Once fully set up, CraftBot launches in browser mode:
CRAFTBOT IS READY — running at localhost:7925 with all 7/8 systems initialized (frontend server, agent backend, MCP servers, skills, integrations, scheduler)
Setup verdict: OpenClaw wins on speed — 60 seconds vs 15-20 minutes. But CraftBot's setup is one-time, and what you get afterward is in a completely different category.
The Interface
Access CraftBot at http://localhost:7925 in your browser.
CraftBot's main interface Chat, Tasks, Dashboard Workspace navigation and the prominent orange Add Living UI button Clean, fast entirely local.
The four sections — Chat, Tasks, Dashboard, Workspace — each serve a distinct purpose. No clutter.
Conversations & Intelligence
Starting simple:
Hello, who are you? → CraftBot: Hey! I'm CraftBot an AI agent built by CraftOS. I can handle pretty much any computer-based task you throw at me from research and coding to file operations, scheduling, and more.
When I asked "What can you do?" — the response listed 7 capability categories:
7 capabilities: Research & Analysis File Management Coding & Automation Scheduling & Monitoring, System Operations, Third-party Integrations and notably Self-improvement (installing new tools and skills at runtime)
That last one Self-improvement is worth pausing on. CraftBot can install new tools and skills while it's running. That's not a chatbot feature. That's an agent feature.
OpenClaw comparison: OpenClaw is reactive — you ask, it responds. CraftBot is designed to be proactive — background tasks, scheduled actions, context awareness across sessions.
Dashboard & Observability
This is where CraftBot separates itself from anything browser-based.
CraftBot's full observability dashboard — Task Statistics, Token Usage breakdown (Input 98%, Output 2%, Cached 14%), System Resources (CPU/memory/disk/thread pool), Usage Patterns with peak hour analytics, MCP server status, Skills (11 enabled), and Model Information showing exactly which model is running
Everything visible, nothing hidden. The Model Information panel shows deepseek/deepseek-v4-flash via openrouter — the exact model, the exact provider. Compare that to cloud tools that hide this behind "proprietary AI."
What the dashboard shows you:
- Task completion rates and failures
- Token consumption broken down by type
- Real-time CPU, memory, disk usage
- Request history and peak usage hours
- Which MCP servers are connected and how many calls they've made
- Exactly which AI model you're running
No browser extension can give you this level of visibility into what your agent is actually doing.
Workspace & File Management
CraftBot has a built-in file browser at localhost:7925/workspace.
CraftBot Workspace — built-in file manager with New Folder, New File, and Upload buttons. The living_ui/ folder is where custom apps built by the agent get stored.
Create folders, create files, upload documents. The agent can read and write to this workspace directly — no manual file sharing required.
MCP Support Explained
I asked CraftBot directly what MCP is:
CraftBot on MCP: It's an open standard that lets AI applications like me connect to external tools and services. Think of it like a universal plugin system — an MCP server runs as a subprocess and exposes tools, resources, and prompts that I can call directly.
This is the architecture that lets CraftBot actually do things — read your Gmail, search GitHub issues, update a Notion page, run a web search. Each MCP server you configure during setup becomes a set of callable tools.
The Tasks & Actions panel on the right shows the agent's current task queue Plan my day and User Profile Interview were active during this session.
Local Model Support (Ollama)
The capability that matters most for privacy: running completely offline.
CraftBot automatically installs Ollama — shown here at 55% progress. The process completes without any manual intervention.
Switch the provider to Local (Ollama) in Settings, choose your model, and CraftBot installs Ollama automatically via winget. Once configured, your data never leaves your machine. No API costs. No internet required.
This is one of the most important features for developers working with sensitive codebases or proprietary information. OpenClaw has no equivalent.
Security Review: How CraftBot's Architecture Handles the 5 Key Tests
Based on my architecture review and documentation analysis CraftBot is designed to pass all 5 key security tests here's why.
Test 1: Read/Write Permissions Access Control
CraftBot's MCP servers can be restricted to read-only or write-only mode at the configuration level. Even if an agent session were somehow compromised, it physically cannot delete or modify files when set to read-only. This is granular permission control at the server layer — not just a UI toggle.
vs OpenClaw: No equivalent granular permission control exists in the browser extension model.
Test 2: Scope Limitation Document Access
CraftBot allows you to define exactly which folders documents or search scopes the agent can access The agent cannot reach anything outside its defined boundary it's enforced at the architecture level You decide what the agent sees.
vs OpenClaw: OpenClaw is browser-scoped by nature CraftBot has local filesystem access but with enforced scope controls a meaningful distinction for sensitive work.
Test 3: Local Data Privacy No Cloud Leakage
CraftBot's core and memory run entirely on local hardware Sensitive data never leaves your machine unless you explicitly connect an external API and you control which APIs are connected With Ollama enabled even the model inference is local zero data transmitted.
Why it matters: For developers working with proprietary code, client data, or regulated industries, this is non-negotiable. Cloud-based agents simply cannot offer this guarantee.
ARCHITECTURE SUPPORTS TEST 4: Approval-Based Execution Human in the Loop
CraftBot's proactive agent can initiate tasks autonomously but it requires explicit user approval before executing any sensitive operation. This human-in-the-loop design means the agent plans and proposes; you authorize and execute No silent background actions on critical operations.
vs OpenClaw: OpenClaw is purely reactive (you command, it responds) CraftBot is proactive but approval-gated a more capable and safer design for agentic workflows.
ENTERPRISE-GRADE SECURITY (TEST 5 PASS): Secret Links & Password Protection MCP Security
Every MCP server connection in CraftBot uses cryptographically secure, unguessable URLs not predictable endpoints that could be discovered or brute-forced. For sensitive deployments OAuth 2.0 password protection can be layered on top adding a second authentication factor to every external integration.
Why it matters: This is enterprise-grade security architecture — the kind you'd expect from production tooling, not an open-source desktop agent. It means CraftBot can be safely used in team or organizational environments where MCP endpoints need to be protected.
Head-to-Head Comparison
| Feature | OpenClaw (Browser) | CraftBot (Desktop) |
|---|---|---|
| Setup time | ~1 minute | ~15-20 minutes |
| Type | Browser extension | Desktop app |
| Interface | Chat | Chat + Dashboard + Workspace |
| File management | ❌ | ✅ Built-in |
| MCP support | Partial | ✅ Native, 10+ servers |
| Local models (Ollama) | ❌ | ✅ Auto-installs |
| Observability dashboard | ❌ | ✅ Full metrics |
| Token usage tracking | ❌ | ✅ Real-time |
| Background/scheduled tasks | ❌ | ✅ |
| Create custom apps (Living UI) | ❌ | ✅ |
| Works offline | ❌ | ✅ (with Ollama) |
| Open source | Partially | ✅ Fully |
Which One Should You Choose?
Choose OpenClaw if:
- You want AI assistance running in under 2 minutes
- Browser-based is sufficient for your workflow
- You're doing simple, reactive tasks
- You want zero terminal interaction
Choose CraftBot if:
- Privacy matters — you want to know exactly what model runs your data
- You need local model support (Ollama) for offline or sensitive work
- You want file management, background tasks, or service integrations
- You want developer-grade observability over your AI agent
- You're building serious agentic workflows
- Open source and self-hosting matter to you
The Bottom Line
These tools aren't competing for the same user.
OpenClaw is a hammer immediate simple good at what it does Great for getting started with AI assistance without friction.
CraftBot is a workshop It costs more setup time demands more from you upfront and has real complexity to navigate But what you get is a local AI agent with observability file management MCP integrations offline model support and a genuine capability ceiling well above any browser extension.
If you're serious about local AI agents and willing to invest the setup time, CraftBot is worth it.
Resources
Have you used OpenClaw, CraftBot, or another local AI agent? What's been your experience especially around privacy and local models?
Drop a comment I read every one. 👇
Disclosure: CraftBot provided access for testing. All opinions and testing are my own.
Top comments (0)