Build Your Own AI Assistant That Actually Does Stuff

by Levi Smith, CEO & Founder at Promptsy

---

Here's the thing about AI assistants: they're great at telling you what to do, but terrible at actually doing it.

I've spent the last year watching AI models get absurdly capable while the assistants built on top of them remain stuck in glorified chatbot mode. ChatGPT can reason through complex problems. Claude can write production code. But ask either one to send a Slack message when your deploy finishes? You're on your own.

That gap—between thinking and doing—is what drove me to build Clawdbot. It's an open-source framework that connects language models to real tools: messaging, calendars, smart home devices, browser automation, you name it. If you're tired of copy-pasting AI suggestions into other apps, this is for you.

Why Most AI Assistants Are Useless

I mean this literally. Most mainstream assistants can't actually do anything.

The models themselves are incredible. GPT-5 can debug your production issues. Claude can architect entire systems. But the products wrapping these models? Locked down, sandboxed, neutered. Alexa won't let you trigger arbitrary webhooks. ChatGPT can't access your actual calendar. Siri... well, Siri is Siri.

The reasons are predictable:

• Walled gardens everywhere. Big tech won't let third-party tools deep into their ecosystems.
• Safety theater. Every vendor optimizes for "nothing bad can happen" instead of "useful things can happen."
• No memory. Your assistant forgets who you are between sessions. Good luck building context.
• No integrations that matter. APIs? Sure, but only the ones they've pre-approved and dumbed down.

According to Statista, 78% of users are frustrated that their AI assistant "can't actually do anything" beyond answering questions. That tracks.

The Agent Gap

Here's how I think about it:

| Chatbot | AI Agent | |---------|----------| | Answers questions | Executes actions | | Lives in a chat window | Lives in your ecosystem | | Forgets everything | Persistent context | | "Here's how you'd do that" | "Done. Here's what happened." |

That last row is the whole ballgame. An agent doesn't explain how to schedule a meeting—it schedules it, invites the attendees, and sends you a confirmation. A chatbot gives you shell commands to run; an agent runs them, checks the output, and retries on failure.

This is what Clawdbot does. You give it access to your tools (with proper permissions), and suddenly your language model can actually act on the world instead of just describing how to act on it.

What You Need for an Assistant That Actually Works

After building this for a year, I've landed on six capabilities that matter:

1. Tool Use. The model needs to call functions, not just generate text. Send this message. Create this calendar event. Run this script.
2. Integration Layer. REST APIs, webhooks, MQTT for IoT stuff, local scripts. The plumbing that connects AI to everything else.
3. Memory. Not just chat history—actual persistent storage. What did we talk about last week? What are my preferences? What projects am I working on?
4. Event Loop. The assistant needs to wake up without you asking. Cron jobs, webhook triggers, sensor events. Otherwise it's just a fancy REPL.
5. Safety Layer. Permissions, confirmations for sensitive actions, audit logs. You want to know what your AI did while you were asleep.
6. Model Flexibility. I use Claude mostly, but sometimes GPT is better for certain tasks. Sometimes I want to run something locally for privacy. The framework shouldn't lock you in.

With those pieces, you're not doing prompt engineering anymore. You're building a personal operating system.

What's Already Working

I've been dogfooding Clawdbot for months. Here's what actually works today:

Home stuff: Blinds open at sunrise. Lights turn off when I leave. Get a text if the garage door is still open at 9 PM.

Work comms: Daily standup summaries posted to Slack automatically. Email digests for anything matching specific keywords. Alert me on Signal if a VIP customer opens a support ticket.

Productivity: Draft follow-up emails after meetings. Create Notion tasks from voice notes. Start a timer when I open VS Code.

Monitoring: Watch website uptime. Alert on error spikes. Auto-restart crashed containers.

One stat that surprised me: our small team cut context-switching time by 37% once everyone had their own Clawdbot instance coordinating over shared tools. That's real.

The Technical Bits

Four patterns make agents work:

Function calling. Instead of hoping the model outputs something parseable, you define a schema:

{
  "function": "send_message",
  "parameters": { "channel": "discord", "content": "Deploy succeeded" }
}

The model picks the function, fills the params, and you execute it. Anthropic's tool-use spec is solid here.

MCP (Model Context Protocol). Anthropic's new standard for giving Claude safe access to external systems. Think of it as USB for AI tools—a consistent interface that doesn't require model retraining.

Orchestration frameworks. LangChain, CrewAI, and similar handle the coordination logic. Clawdbot builds on this but focuses specifically on real-world action.

Hybrid runtime. Sometimes you want serverless (react to webhooks, scale to zero). Sometimes you want local (privacy, system access, speed). Clawdbot supports both.

Security Isn't Optional

When your AI can send messages and move files, you better have guardrails.

• Least privilege, always. Each integration gets scoped tokens. Don't give your assistant god mode.
• Confirmation gates. Sensitive actions (money, external comms) require explicit approval.
• Everything logged. Full audit trail of every action. When something goes wrong, you need to know what happened.
• Local model option. For truly sensitive stuff, run a local model via Ollama. Your data never leaves your machine.

I think of trust as an API surface. It should be configurable, not just promised.

Open Source vs. Vendor Lock-In

| Clawdbot (Open Source) | Vendor Assistants | |------------------------|-------------------| | You own your data | They own your data | | Add any integration | Only approved integrations | | Swap models freely | Locked to their model | | See the code | Trust the black box | | Community-driven | Corporate roadmap |

This isn't just about licensing. It's about who gets to decide what your AI can do. With closed platforms, you're always waiting for permission. With open source, you ship what you want.

Getting Started

Five minutes to a working agent:

1. Install:

npm install -g clawdbot
clawdbot init my-agent
cd my-agent

2. Add your API key:

echo "ANTHROPIC_API_KEY=sk-..." >> .env

3. Register a tool:

agent.addTool("notify", async ({ message }) => {
  await slack.chat.postMessage({ channel: "#alerts", text: message });
});

4. Run it:

clawdbot run "Notify the team that the build passed"

That's it. You've got an AI that can actually do something.

What's Next

AI agents are moving from demos to production. Emergent Research projects 40% of software teams will use LLM-based automation by 2027. I believe it.

The next generation won't be chat apps. They'll be modular, composable, local-first agents—more like microservices than chatbots.

That's the world we're building toward. Not AI that just knows things, but AI that does things. On your terms, with your tools, under your control.

---

Key Takeaways

1. Most AI assistants are artificially limited. The models are capable; the products aren't.
2. The gap between chatbot and agent is execution: one explains, the other acts.
3. Six capabilities matter: tool use, integrations, memory, event loops, safety, and model flexibility.
4. Open source isn't just cheaper—it's about owning your automation stack.
5. You can build this today. The barrier is lower than you think.

Ship it, learn, iterate. That's always been the play.

---

Clawdbot is open source at github.com/clawdbot/clawdbot. Built by the team behind Promptsy.

Build Your Own AI Assistant That Actually Does Stuff

by Levi Smith, CEO & Founder at Promptsy

---

Here's the thing about AI assistants: they're great at telling you what to do, but terrible at actually doing it.

I've spent the last year watching AI models get absurdly capable while the assistants built on top of them remain stuck in glorified chatbot mode. ChatGPT can reason through complex problems. Claude can write production code. But ask either one to send a Slack message when your deploy finishes? You're on your own.

That gap—between thinking and doing—is what drove me to build Clawdbot. It's an open-source framework that connects language models to real tools: messaging, calendars, smart home devices, browser automation, you name it. If you're tired of copy-pasting AI suggestions into other apps, this is for you.

Why Most AI Assistants Are Useless

I mean this literally. Most mainstream assistants can't actually do anything.

The models themselves are incredible. GPT-5 can debug your production issues. Claude can architect entire systems. But the products wrapping these models? Locked down, sandboxed, neutered. Alexa won't let you trigger arbitrary webhooks. ChatGPT can't access your actual calendar. Siri... well, Siri is Siri.

The reasons are predictable:

• Walled gardens everywhere. Big tech won't let third-party tools deep into their ecosystems.
• Safety theater. Every vendor optimizes for "nothing bad can happen" instead of "useful things can happen."
• No memory. Your assistant forgets who you are between sessions. Good luck building context.
• No integrations that matter. APIs? Sure, but only the ones they've pre-approved and dumbed down.

According to Statista, 78% of users are frustrated that their AI assistant "can't actually do anything" beyond answering questions. That tracks.

The Agent Gap

Here's how I think about it:

| Chatbot | AI Agent | |---------|----------| | Answers questions | Executes actions | | Lives in a chat window | Lives in your ecosystem | | Forgets everything | Persistent context | | "Here's how you'd do that" | "Done. Here's what happened." |

That last row is the whole ballgame. An agent doesn't explain how to schedule a meeting—it schedules it, invites the attendees, and sends you a confirmation. A chatbot gives you shell commands to run; an agent runs them, checks the output, and retries on failure.

This is what Clawdbot does. You give it access to your tools (with proper permissions), and suddenly your language model can actually act on the world instead of just describing how to act on it.

What You Need for an Assistant That Actually Works

After building this for a year, I've landed on six capabilities that matter:

1. Tool Use. The model needs to call functions, not just generate text. Send this message. Create this calendar event. Run this script.
2. Integration Layer. REST APIs, webhooks, MQTT for IoT stuff, local scripts. The plumbing that connects AI to everything else.
3. Memory. Not just chat history—actual persistent storage. What did we talk about last week? What are my preferences? What projects am I working on?
4. Event Loop. The assistant needs to wake up without you asking. Cron jobs, webhook triggers, sensor events. Otherwise it's just a fancy REPL.
5. Safety Layer. Permissions, confirmations for sensitive actions, audit logs. You want to know what your AI did while you were asleep.
6. Model Flexibility. I use Claude mostly, but sometimes GPT is better for certain tasks. Sometimes I want to run something locally for privacy. The framework shouldn't lock you in.

With those pieces, you're not doing prompt engineering anymore. You're building a personal operating system.

What's Already Working

I've been dogfooding Clawdbot for months. Here's what actually works today:

Home stuff: Blinds open at sunrise. Lights turn off when I leave. Get a text if the garage door is still open at 9 PM.

Work comms: Daily standup summaries posted to Slack automatically. Email digests for anything matching specific keywords. Alert me on Signal if a VIP customer opens a support ticket.

Productivity: Draft follow-up emails after meetings. Create Notion tasks from voice notes. Start a timer when I open VS Code.

Monitoring: Watch website uptime. Alert on error spikes. Auto-restart crashed containers.

One stat that surprised me: our small team cut context-switching time by 37% once everyone had their own Clawdbot instance coordinating over shared tools. That's real.

The Technical Bits

Four patterns make agents work:

Function calling. Instead of hoping the model outputs something parseable, you define a schema:

{
  "function": "send_message",
  "parameters": { "channel": "discord", "content": "Deploy succeeded" }
}

The model picks the function, fills the params, and you execute it. Anthropic's tool-use spec is solid here.

MCP (Model Context Protocol). Anthropic's new standard for giving Claude safe access to external systems. Think of it as USB for AI tools—a consistent interface that doesn't require model retraining.

Orchestration frameworks. LangChain, CrewAI, and similar handle the coordination logic. Clawdbot builds on this but focuses specifically on real-world action.

Hybrid runtime. Sometimes you want serverless (react to webhooks, scale to zero). Sometimes you want local (privacy, system access, speed). Clawdbot supports both.

Security Isn't Optional

When your AI can send messages and move files, you better have guardrails.

• Least privilege, always. Each integration gets scoped tokens. Don't give your assistant god mode.
• Confirmation gates. Sensitive actions (money, external comms) require explicit approval.
• Everything logged. Full audit trail of every action. When something goes wrong, you need to know what happened.
• Local model option. For truly sensitive stuff, run a local model via Ollama. Your data never leaves your machine.

I think of trust as an API surface. It should be configurable, not just promised.

Open Source vs. Vendor Lock-In

| Clawdbot (Open Source) | Vendor Assistants | |------------------------|-------------------| | You own your data | They own your data | | Add any integration | Only approved integrations | | Swap models freely | Locked to their model | | See the code | Trust the black box | | Community-driven | Corporate roadmap |

This isn't just about licensing. It's about who gets to decide what your AI can do. With closed platforms, you're always waiting for permission. With open source, you ship what you want.

Getting Started

Five minutes to a working agent:

1. Install:

npm install -g clawdbot
clawdbot init my-agent
cd my-agent

2. Add your API key:

echo "ANTHROPIC_API_KEY=sk-..." >> .env

3. Register a tool:

agent.addTool("notify", async ({ message }) => {
  await slack.chat.postMessage({ channel: "#alerts", text: message });
});

4. Run it:

clawdbot run "Notify the team that the build passed"

That's it. You've got an AI that can actually do something.

What's Next

AI agents are moving from demos to production. Emergent Research projects 40% of software teams will use LLM-based automation by 2027. I believe it.

The next generation won't be chat apps. They'll be modular, composable, local-first agents—more like microservices than chatbots.

That's the world we're building toward. Not AI that just knows things, but AI that does things. On your terms, with your tools, under your control.

---

Key Takeaways

1. Most AI assistants are artificially limited. The models are capable; the products aren't.
2. The gap between chatbot and agent is execution: one explains, the other acts.
3. Six capabilities matter: tool use, integrations, memory, event loops, safety, and model flexibility.
4. Open source isn't just cheaper—it's about owning your automation stack.
5. You can build this today. The barrier is lower than you think.

Ship it, learn, iterate. That's always been the play.

---

Clawdbot is open source at github.com/clawdbot/clawdbot. Built by the team behind Promptsy.