Claude is still the AI tool I reach for most days and that's probably not going to change any time soon. But even on the Pro plan, I still hit the limit caps, especially when I've been doing visual work with Artifacts or Claude Design, and then there isn't much left for regular chat without turning on extra usage. Opus eats through that allowance way faster than Sonnet, and I run Opus a lot. There's also the privacy side - everything I send Claude lives on Anthropic's servers, and while that's fine for most things, the personal documents I'd rather keep on my device. Plus, having a tool that works regardless of whether Anthropic's having a bad server day is its own kind of insurance.

A local LLM on its own doesn't fully cover what Claude does though. The models themselves are solid - mine handles most of my day-to-day prompting without much complaint. What's missing is the workspace side, such as projects with persistent context and a knowledge base. So if I wanted something that vaguely felt more like a productivity space within a chatbot, I'd have to create it myself, and that's where Obsidian comes in.

Claude Pro is near-perfect, but it's good to have options

The usage caps are where it starts to sting

The weekly cap honestly hasn't been a huge issue for me outside of design work, where Claude burns through it faster than anywhere else. It's the 5-hour rolling window that gets me. Anthropic doesn't publish exact numbers but Opus burns through tokens roughly 3-5x faster than Sonnet, and I run Opus on high reasoning effort for most of the research and studying I do with Claude. So the math doesn't work in my favour, and I hit that 5-hour wall way earlier than the 45-ish messages estimate Anthropic floats. When that happens I'm locked out of the session unless I enable Extra Usage and pay more.

The goal wasn't really to replace Claude though. It was more about figuring out which chunk of my Claude usage could happen somewhere else so the 5-hour allowance could go toward the work that actually needs the horsepower. Day-to-day brainstorming and notes-adjacent stuff doesn't fall into that category.

Obsidian came into the picture for a slightly unintuitive reason. It's a notes app rather than an AI tool, and that's actually why it works for this. The vault is local and my notes already live there, so anything I layer on top of it has a context layer already built in. Folder structures behave kind of like Projects (in terms of organization, not memory), so you can point an AI model you bring into Obsidian at a specific directory and it has that scope to work with.

Setting up Obsidian with my local LLM

It's dead simple actually

I've set up Obsidian with a local LLM before through the Copilot plugin, but that was back when I was still a bit newer to local models and defaulted to gpt-oss20B. Now, I've been running better models such as Qwen 3.5 9B and Gemma 4 E4B, and have also tested various local LLM runners beyond LM Studio. But I still did it through LM Studio because Copilot's provider list is slightly limited, and it hooked up cleanly before so I didn't see a reason to try and mess with the setup.

If you want to set up your model with Obsidian like I did, first, download your pick in LM Studio, go to the Dev tab, and enable CORS in the Server Settings. Also, copy the exact model name while you're at it. Then in Obsidian, ensure the Copilot plugin is enabled, open the settings, and head to Models to add your model. Paste the model name in the name field, set your provider, put http://localhost:1234/v1 in the Base URL, leave the API field empty, and turn on your preferred model capabilities (CORS must be enabled). And that's all there is to it, now your model is in Obsidian.

The more interesting part is what you do with your model after the fact. Copilot does have its own version of projects, but that's for the paid tier, and I wanted to keep this free and local. Custom prompts handle the system prompt side, they're simply reusable .md files in the copilot-custom-prompts folder, triggered via slash command. This is as close as you'll get to "persistent context" in the free setup, but it works. I have multiple for different tasks, like research and brainstorming for my novel.

For the knowledge base side, typing @ in chat lets me point Copilot at a specific folder, which becomes the scope for that session. And the [[note links]] syntax pulls individual notes into chat the same way it works everywhere else in Obsidian. Chat history can also be saved as a regular note in the vault, which means conversations live in the same place as everything else.

Where this combo actually lands for me

It's a solution, but not to everything

It's not Projects or memory, but it's a usable system in the case of Claude being unavailable to me due to usage caps, server shutdown, or wifi outage. Opus 4.7 chews through long-context reasoning in a way a 4.5B model just can't, which is what you'd expect when one model runs on Anthropic's servers and the other runs on my GPU. But Opus is overkill for day-to-day chatting anyway, a local model handles all the basic stuff.

There's also stuff this combo does that Claude can't. Output lands directly in my notes - no copy-paste cycle - and conversations save as markdown files in the vault, accessible right alongside everything else. Privacy is the other obvious reason to go local. I have a million health reports on my drive and don't actually know how to properly understand them, so handling that locally, in a contained and organized space, is a major win.

Claude still gets the heavy stuff

So a few months in, I think this combo is doing what I needed it to. Claude's still where the harder work happens, while the smaller daily stuff lives in Obsidian now with my local LLM layered over it, which has freed up the Pro cap for the deeper work revolving around designing. It's definitely not a replacement, but I think it's pretty neat and usable.