A glowing interface links digital files and settings to a secure computer, cloud storage, and Windows/Linux systems.
Anyone building a "second brain" in Obsidian eventually slams into the same uncomfortable crossroads: cloud AI models are fast, clever, and effortless until you remember you are shipping your private journals to someone else's servers, while local models keep your secrets on your own SSD at the cost of turning a modest laptop fan into a leaf blower. In a recent guide published by It's FOSS, Linux tinkerer Sreenath laid out a clean workaround—a hybrid AI setup inside Obsidian that routes sensitive personal notes to a local model powered by Ollama while sending everyday work and learning queries to Google's Gemini API.

It is an appealing blueprint for Windows and Linux power users alike, and the core idea holds up remarkably well. Still, if you try to follow a generic walkthrough word for word right now, two things will trip you up: Copilot for Obsidian has undergone a major V4 settings overhaul that moved the menus around, and "free" or "local" AI carries finer privacy and hardware print than most tutorials spell out. Here is how the hybrid architecture actually works, how to deploy it cleanly on Windows as well as Linux, and where the real boundaries sit.


A glowing interface links digital files and settings to a secure computer, cloud storage, and Windows/Linux systems. The Hybrid Vault Strategy: Separating Privacy From Speed​

The philosophy behind Sreenath's It's FOSS workflow is straightforward: stop forcing one model to do every job. Instead of choosing between absolute isolation and cloud convenience, you split your knowledge base by sensitivity—even maintaining separate Obsidian vaults for personal journaling and work or study material if you want hard separation at the folder level.

In practice, that means:

  • Local AI (Ollama + Llama 3.2 3B) handles private diaries, financial notes, health logs, and anything else you never want leaving your PC.
  • Cloud AI (Google Gemini via API) handles coding notes, study outlines, technical documentation, and brainstorming where model intelligence and response speed matter more than strict secrecy.

What "Local" and "Free" Actually Mean for Privacy​

Before moving a single note, it pays to clear up two common misconceptions that can catch privacy-conscious Windows users off guard:

  1. A local LLM does not sandbox the plugin. Obsidian runs community plugins in Restricted Mode by default precisely because third-party extensions execute arbitrary JavaScript on your machine with broad access to local files and network connections. Obsidian performs automated security scans and assigns safety scorecards to plugins in its official directory, and Copilot's own documentation notes that local models keep inference traffic on your hardware. Even so, Copilot V4's settings guide warns that agent modes and optional skills—such as web search or web fetch—can still reach out to external network services when triggered. Always check which provider, agent, and skills are active before feeding confidential files into a chat pane.
  2. "Gemini is free" comes with data-use strings attached. The It's FOSS walkthrough notes using Gemini largely because it costs nothing to get started. According to Google AI Studio's Gemini API pricing documentation, Google's Free Tier indeed waives token charges for supported models, but explicitly states that Free Tier prompts and content are used to improve Google's products. By contrast, Google's Paid Tier states that submitted content is not used for product improvement, and also unlocks models unavailable on the Free Tier (such as Gemini 4 Carbon). If your "work" notes include proprietary company code or client data, a free-tier API key is not a compliance shield.
AI RouteWhere Prompts GoUsed for Provider Training?Hardware Load on PCOffline Support
Ollama (Local Model)Stays on localhost:11434NoHigh (VRAM / RAM / CPU)Full
Google AI Studio (Free Tier)Google cloud serversYes, per Google's Free Tier termsMinimalNone
Google AI Studio (Paid Tier)Google cloud serversNo, per Google's Paid Tier termsMinimalNone

Section Summary: Splitting vaults or workflows between local and cloud models gives you practical control over speed versus confidentiality, provided you account for third-party plugin permissions and free-tier cloud data policies.


Copilot for Obsidian: Installing It and Navigating the V4 Shift​

Despite the name, Copilot for Obsidian is an independent community plugin with zero affiliation with Microsoft Copilot. It serves as the traffic controller inside Obsidian, letting you switch between local endpoints and cloud APIs from one sidebar without paying for a Copilot subscription—Copilot's documentation confirms that Agent Chat, Quick Chat, Quick Ask, BYOK ("Bring Your Own Key"), and local model support are all included on its Free tier.

Installing the Extension Safely​

  1. Open Obsidian Settings with Ctrl +, and select Community plugins in the left sidebar.
  2. Turn off Restricted Mode (select Turn on community plugins).
  3. Click Browse, search for Copilot, select Install, and click Enable.
  4. Remember that Obsidian intentionally disables automatic updates for community plugins as a security safeguard. Periodically revisit Settings → Community plugins → Check for updates to pull security patches and bug fixes.

Bridging the V3 vs. V4 Menu Gap​

If you read older guides—including the menu labels shown in the It's FOSS tutorial—you will see instructions pointing to Basic → Set Keys and Model → Add Model. In Copilot V4, the developer rebuilt the plugin around Agent Chat on desktop (powered by OpenCode, Claude Code, or Codex) alongside lightweight Quick Chat and Quick Ask, and consolidated external credentials under a dedicated BYOK tab.

Three V4 details matter immensely if you use Obsidian across multiple Windows or Linux PCs:

  • Keys do not sync with your vault: Copilot stores provider API keys in your device's Obsidian Keychain, not in the vault's data.json file. Syncing a vault via OneDrive, Dropbox, or Obsidian Sync will not carry your API keys to another computer; you must enter them separately on each machine.
  • BYOK feeds Quick Chat and OpenCode, not Claude/Codex agents: Entering an API key under BYOK powers Quick Chat and the OpenCode agent. The Claude and Codex agent tabs rely on your existing Claude Code CLI or Codex CLI subscription logins instead.
  • First-time Agent setup: Under Settings → Copilot → Basic → Agents, the OpenCode tab offers a Managed by Copilot option with a Download & install button so you do not have to wrangle binaries manually.

Connecting Cloud AI: Setting Up Google Gemini​

With Copilot installed, wiring in Google Gemini takes less than five minutes:

  1. Visit Google AI Studio, click Create API key, assign it to a new or existing project, and copy the generated key.
  2. In Obsidian, press Ctrl +, and open Copilot in the plugin settings list.
  3. On Copilot V4, open the BYOK tab, click Add a provider, select Google (or Gemini), and paste your API key.
  4. Choose your target model from the provider's model list—Sreenath used Gemini 2.5 Flash for fast, everyday note work, though available model lineups evolve over time, so treat the live picker in settings as your source of truth. Click Test and then Save. (If you are running an older Copilot V3 build, look under Basic → Set Keys instead.)
  5. Open Basic → Agents (and Quick Chat or OpenCode) to verify the model is enabled and set as your default where desired.

Once saved, Copilot verifies your provider status automatically whenever you revisit the BYOK tab and displays badges if a key goes invalid or hits a connection error.


Setting Up Local AI With Ollama (Windows and Linux)​

Where the It's FOSS tutorial focuses on Linux shell commands, Windows PCs actually have an even smoother on-ramp—along with a few storage and VRAM traps worth dodging upfront.

Step 1: Install Ollama on Your OS​

  • On Linux: Run Ollama's official install script in a terminal:
    curl -fsSL [url]https://ollama.com/install.sh[/url] | sh
  • On Windows (Windows 10 22H2+ or Windows 11): Download the native installer (OllamaSetup.exe) from Ollama's site or run the official PowerShell command:
    irm [url]https://ollama.com/install.ps1[/url] | iex

Ollama's Windows documentation notes that the desktop installer does not require Administrator rights and installs into your user profile directory by default, requiring at least 4 GB for the binary before you even download a model. It runs quietly in the system tray, auto-updates via the taskbar menu, and supports NVIDIA GPUs (driver 551.61 or newer), AMD Radeon cards via ROCm v7/HIP7, and an automatic Vulkan fallback for broader Windows GPU support.

Windows Pro Tip: If your C:\ drive is tight on space, set a user environment variable named OLLAMA_MODELS pointing to a secondary SSD (for example, D:\OllamaModels) before pulling large weights, or pass /DIR="D:\Ollama" to OllamaSetup.exe during installation.

Step 2: Verify the Local Server and Pull llama3.2:3b​

Open a browser and navigate to [url]http://localhost:11434[/url]. You should see the plain-text confirmation: "Ollama is running".

Next, open Command Prompt, PowerShell, or your Linux terminal and pull the 3-billion-parameter Llama 3.2 model:

ollama run llama3.2:3b

Ollama's library lists llama3.2:3b at roughly 2.0 GB, making it light enough to stay snappy on modest GPUs while still remaining tolerable for occasional CPU-only testing. Once it downloads and drops you into an interactive prompt, type /bye to exit, then run:

ollama list

Copy the exact model name shown (llama3.2:3b), and if you want to double-check that your GPU is actually doing the heavy lifting during a query, run ollama ps in another terminal window. Look at the PROCESSOR column: 100% GPU means everything fit in VRAM; if you see 100% CPU or a split percentage, your GPU ran out of memory and spilled over to system RAM.

You should also keep an eye on context length. While Llama 3.2 supports up to a 128K context window on paper, Ollama's official context documentation notes that it defaults to a 4K (4,096-token) context window on GPUs with under 24 GiB of VRAM (scaling to 32K for 24–48 GiB VRAM) to avoid choking memory. If you feed long notes into local chat and notice the model "forgetting" the top of the page, adjust the context slider in Ollama's Windows desktop settings or set the OLLAMA_CONTEXT_LENGTH environment variable—just keep an eye on ollama ps afterward to make sure you haven't pushed the model off your GPU.

Step 3: Wire Ollama Into Copilot​

  1. In Obsidian, open Settings → Copilot → BYOK → Add a provider.
  2. Under the Self-host group, choose Ollama (no paid Self-Host license is required).
  3. Verify the Base URL is populated as [url]http://localhost:11434/v1[/url] and leave the API key blank (it is optional for local Ollama).
  4. Enter or select llama3.2:3b as the model ID, click Test, and click Save. (In older Copilot V3 setups, this lives under Model → Add Model with the Base URL left at its default.)
  5. If Test passes on Windows but Quick Chat refuses to stream replies, edit the Ollama provider entry and toggle Enable CORS—responses will appear all at once upon completion rather than token by token, bypassing local browser-style CORS blocks.

Section Summary: Native Windows installation, OLLAMA_MODELS drive redirection, ollama ps VRAM checks, and the V4 BYOK endpoint ([url]http://localhost:11434/v1[/url]) turn a Linux-centric guide into a rock-solid local AI setup on Windows 10 and 11.


Day-to-Day Workflow: Context, Vision Limits, and Chat Backups​

Once both providers are wired in, switching between local privacy and cloud horsepower happens right inside Copilot's sidebar—giving you four everyday tools that make Obsidian far more capable than a plain chat window:

  • Multi-Note @ Context: By default, Copilot attaches your currently open note to new chats. Typing @ inside the chat box lets you pull in multiple related notes, folders, or selections so the model can synthesize ideas across your vault rather than guessing in a vacuum.
  • Screenshots in Chat (With a Crucial Model Catch): The It's FOSS guide highlights pasting screenshots of config errors straight into the sidebar to ask what went wrong. Here is the catch: toggling "Vision" in a plugin dialog does not give a text-only model eyesight. Ollama's library explicitly classifies llama3.2:3b as a text-in / text-out model. If you want to debug screenshots, either route non-sensitive captures to a multimodal cloud model like Gemini or pull a dedicated local vision model in Ollama.
  • Insert or Replace at Cursor: Instead of playing clipboard ping-pong, Quick Ask and Copilot's sidebar responses provide Insert and Replace actions that drop formatted Markdown answers directly at your active cursor position inside the note.
  • Markdown Chat Logs and Backlinks: Copilot archives conversations inside your vault under copilot/copilot-conversations as standard Markdown files—and in V4, stores attached images once per conversation under copilot-conversations/attachments/—while linking back to the active note. Because those chats remain plain Markdown, they stay searchable and future-proof even if you uninstall the plugin years from now.

One extra Windows wrinkle to note if you dive deeper into Copilot V4's Skills system: Copilot's documentation warns that creating shared filesystem links across agents on Windows may require either Administrator privileges or turning on Settings → Privacy & security → For developers → Developer Mode, and keeping your vault inside OneDrive or Dropbox can cause sync clients to break those links. If you only want Quick Chat and basic OpenCode Q&A, you can ignore Skills altogether.


The Bottom Line: Let AI Format, Not Think for You​

The smartest takeaway from It's FOSS's hybrid setup isn't a clever plugin trick—it is restraint. Local and cloud LLMs alike still hallucinate with complete confidence, so outsourcing primary research or core learning to an Obsidian sidebar is asking for poisoned notes. Where this dual-engine system shines is friction removal: using Gemini on non-sensitive material to turn messy meeting scraps into clean Markdown tables, and leaning on a local Ollama model to summarize journal entries or brainstorm private outlines without a single byte leaving your PC.

Have you split your Obsidian vaults between local Ollama weights and cloud APIs on Windows 11? Drop your GPU specs, favorite local model sizes, and any V4 BYOK quirks you've run into in the WindowsForum discussions—especially if you've found a sweet spot for VRAM versus context length.

 

References

  1. My Hybrid AI System in Obsidian: Local and Cloud AI Integration - It's FOSS It's FOSS 2026-11-22T05:18:03+00:00
  2. Download Ollama on Linux ollama.com
  3. obsidian.md obsidian.md