My Journey to Building Benny Chat
When the AI Tools Available Just Aren't Enough
I've been on a winding path through knowledge management and AI tools, and today I want to share how I ended up building my own solution. If you caught my video a few months ago about using Obsidian with Claude, you might remember how excited I was about connecting these tools through MCPs (Model Context Protocol). The response was fantastic, but I've since moved on to something even better.
Let me introduce you to Benny Chat – my custom-built replacement for the Claude-Obsidian workflow. While I'll share how to access Benny at the end of this post, I first want to walk you through the journey that led me here. It's been quite the rabbit hole!
If you’d rather watch than read, there’s a companion video on our YouTube channel that walks through the same material step by step—feel free to pause here and check it out.
The Knowledge Management Quest
My journey began with Roam Research, a tool that excels at connecting notes by making each note a referenceable object (similar to website URLs). This creates webs of connections between different notes in a shareable format. While I loved the connectivity, I struggled with its node-based structure – each block of text is a specific node. This just didn't match how I think and write.
This frustration led me to build my own application (not my first rodeo). I created "D&D Notes" where I could add entries like "The Dragon's Castle" and tag them as places, people, organizations, etc. Notes could reference other notes, creating that connectability I wanted.
But then I discovered Obsidian, which is essentially Roam Research for documents, but with much more flexibility. You can shape documents however you want without the strict structure. I switched immediately, pulled all my Roam Research content over, and moved forward.
What really captivated me about Obsidian was its graph view – the ability to visualize connections between notes. I found myself regularly checking this graph, watching how my note structure evolved over time. Plus, the plugin ecosystem allowed me to extend and customize Obsidian to fit my specific needs.
Adding AI to the Equation
As I dove deeper into AI, I explored tools like Obsidian Smart Connections, which adds AI capabilities to connections between notes. The semantic similarity search was particularly interesting – instead of searching for exact text matches, I could find notes with similar concepts. Looking for "Hero's Fortress" might lead me to "Mario's Castle" through semantic connections.
Pretty neat! But there was a significant problem – keeping Smart Connections up to date was a hassle. The embedding models changed frequently, requiring me to re-embed everything. Every time I opened the app, I'd wait a long time for embeddings to update before I could use the fuzzy search feature I actually wanted.
I also didn't love the chat interface. Obsidian Copilot offered a prettier alternative, but after my experience with Smart Connections, I was hesitant. I realized I didn't actually want to chat with my notes – I wanted to reference them in a robust system that could take action.
At this point, I’d pared everything back to core Obsidian and a couple of must-have plugins. The Obsidian Text Generator covered quick drafting and rewrites, which felt “good enough” for daily writing—at least until I saw how AI could tackle tasks.
The Task Management Revelation
Everything changed when I saw a post from Austin, the founder of Saner. He demonstrated breaking down a task into pieces, prioritizing them, and assigning due dates – all through a chat interface. I was immediately intrigued.
While Saner had good connections, I got stuck in the onboarding process. It pushed too much at once and didn't quite understand my context when I tried to use the chat as I'd seen in Austin's demo.
But the idea of AI managing tasks stuck with me. I explored Tana, which bills itself as an "AI native workspace" using "super tags" to organize notes. While interesting, I found its AI capabilities frustratingly limited:
Me: What are my current goals?
Tana: I'm sorry, but I don't have access to your personal information or current goals.
What gives? For an AI task workspace, this was a direct violation of what I expected. Even when I pointed it to specific tasks in my workspace, it couldn't access them. The gap between the AI capabilities and the actual tool was maddening.
The MCP Breakthrough
Then I discovered MCPs (Model Context Protocol) and had a revelation. I could use Claude Desktop with Todoist through MCPs to create what I wanted:
Me: Hey, good morning, what tasks are on my to-do list?
Claude Desktop: Good morning! Here are your tasks for today: - Draft the product update email (Due: today) - Review pull request #42 in GitHub (Due: today) - Schedule meeting with design team (Due: tomorrow) - Follow up on client feedback (Due: tomorrow) Let me know if you want to mark any as complete, add a new task, or get more details about any item!
And it would pull down my tasks! I could ask for overdue tasks, create new ones – everything I wanted from Tana but couldn't get. This was ideal.
I had this big realization: MCPs completely replaced a major company’s task management software. There was no turning back.
But there was one significant problem – Claude's uptime is questionable. With the entire world hitting their system, they don't maintain 99.99% uptime. There were times I simply couldn't use Claude throughout the day. Other MCP systems weren't as reliable, and since I was running MCPs on my local machine, the servers went down frequently. I was constantly restarting Claude to get them back up.
I had this amazing system with brutally bad uptime – what felt like 80% at best. That's when I decided: enough. Time to build my own solution.
Knowledge Management vs. Task Tracking
Before building, I needed to clarify the distinction between knowledge management and task tracking in my workflow:
Knowledge Management:
Long-term documents written by me
I control what goes in (no AI auto-importing)
Location and structure are bespoke
Used for brainstorming and organizing ideas
Task Tracking:
Short-term tasks and engineering logs
History matters but is temporary
Feeds into long-term documents
Action-oriented and prioritized
My AI needs differ between these spaces. For task tracking, I'm fine with AI-written artifacts – the AI can write the entire task description based on my input. I love auto-tagging and filing for tasks but am more cautious about it for knowledge management. Semantic search is essential for knowledge management but less critical for well-structured tasks.
This clarity helped me concentrate on what I really needed to build.
Introducing Benny Chat
So I built Benny Chat – a custom solution focused on task tracking while connecting to my knowledge management system. Here's what it can do:
1. Task Management: I can ask "What are my current to-dos for today?" and Benny pulls them from Todoist. I can create new tasks, and they appear with checkboxes that maintain context across different chats.
2. Knowledge Access: I can search my Obsidian notes with queries like "Look up my notes related to React Native" and get relevant results. Benny can display full notes with metadata and AI summaries.
3. Visual Content: It even renders my Excalidraw files directly in the chat, allowing me to reference visual notes and diagrams.
4. Web Search: Through Perplexity integration, I can perform web searches without leaving the chat.
All of this runs on a server, so I can use it on my phone or on the go. I've set it up with connected all the necessary MCPs (I even wrote a custom GitHub search MCP for my specific needs).
The result is better than Claude Desktop + MCPs for my specific needs – more reliable, more tailored to my workflow, and more integrated with my existing tools.
Try Benny Chat Yourself
If this sounds useful, here’s how you can start experimenting with Benny Chat:
Join our Stable Discussion AI Discord community where you'll connect with AI experts and get early access to tools like Benny.
Subscribe to our paid plan on Stable Discussion, which also includes access to our Discord community.
Both of these options options will give you access to Benny Chat and other AI prototype tools we're developing. You'll be able to use Todoist, GitHub repos as a source (great for Obsidian notes), and Perplexity for web searches.
I'm excited to see what you make of Benny Chat and how it might transform your workflow the way it has mine. After this exploration, I've found that sometimes the best solution is the one you build yourself – especially when it comes to integrating AI with your personal knowledge and task systems.
What knowledge management and task tracking tools are you using? Have you tried connecting them with AI? I'd appreciate hearing about your experiences in the comments!