Obsidian Smart Compose AI Plugin (2026): Local-First AI With Real Trade-Offs

A community-built AI plugin for Obsidian that takes privacy seriously. The AI quality is OK; the philosophy is the actual feature.

Overview

Obsidian is a markdown-first knowledge base app, popular with users who specifically want their notes stored as plain files on local disk rather than in a SaaS vendor’s database. Obsidian itself is free for personal use, has a paid commercial license for businesses, and supports a vibrant community-plugin ecosystem.

Smart Compose is the leading AI plugin in that ecosystem. It supports both cloud-API backends (OpenAI, Anthropic, Google’s Gemini API) where you bring your own key, and local-model backends (via Ollama, llama.cpp, or LM Studio) where models run entirely on your machine. The local-model support is the headline differentiator — for users who specifically want AI inside their notes without sending those notes to a cloud vendor, this is the cleanest implementation we’ve found.

This review covers Smart Compose v2.4.1 on Obsidian 1.7.x, tested over ten weeks with Ollama (running Llama 3.2 70B and Mistral Large locally on an M3 Max with 64GB), the OpenAI API (GPT-5), and the Anthropic API (Claude Opus 4.7). The review focuses on whether the plugin earns its place — and on the privacy proposition that justifies it.

How we tested

We installed Smart Compose into a working Obsidian vault containing roughly 1,200 notes accumulated over several years. Across ten weeks we used it for the standard AI-in-notes workflows: writing assistance (drafting, editing), in-note Q&A (asking questions about content already in the vault), summarization (long notes into shorter ones), and tag-and-link suggestions (the AI suggests where notes should connect into the graph).

We tested all three configured backends:

• Local via Ollama: Llama 3.2 70B-Instruct (40GB on disk, fits in 64GB unified memory), Mistral Large 24B-Instruct (lighter alternative, faster but lower quality).
• OpenAI API: GPT-5 with our own API key, billed token-metered.
• Anthropic API: Claude Opus 4.7 with our own API key.

For comparison context, we ran a parallel evaluation of Notion AI 3 (reviewed separately) on a parallel small Notion workspace.

What works

Local-model support genuinely works. This is the feature that justifies the plugin’s existence. With Ollama installed and Llama 3.2 70B downloaded, Smart Compose configured cleanly via the plugin settings (we pointed it at localhost:11434 and selected the model). All AI requests during local-model testing stayed on the local machine. We verified this by monitoring outbound network traffic during plugin invocations; nothing left the laptop. For users with privacy concerns about sending their notes to cloud vendors, this is the real product.

Cloud-API backends use your own key. When we configured OpenAI or Anthropic APIs, the plugin used our own API keys and our data went to those vendors directly under their respective API privacy policies (which for both vendors mean: no training on data, but standard cloud-vendor exposure). This is meaningfully different from Notion AI or ChatGPT — your billing relationship and your data exposure are with the model provider, not with the plugin.

Open-source plugin code. Smart Compose is open-source on GitHub. We read parts of the codebase (the API routing layer, the context-assembly logic) and the architecture is sensible, the dependencies are reasonable, and the security posture appears sound. For users who want to verify what the plugin does with their data, the option to audit is available.

Note-graph integration. The plugin can use Obsidian’s wiki-link graph as context, pulling in linked notes and tagged content when generating completions. This is a real Obsidian-specific advantage — the AI sees your knowledge structure in a way generic AI tools don’t.

No subscription overhead. The plugin itself is free. Cloud API costs are pay-as-you-go and you control the bill. Local-model use is free (after the upfront hardware cost). For users tired of stacked SaaS subscriptions, the cost model is friendly.

Where it falls short

Hardware requirements for local models. A 70B model is the smallest that produces output quality comparable to (but still below) cloud frontier models. Running 70B requires 64GB+ unified memory on a Mac (M2/M3 Ultra or M4 Max 64GB+) or a 24GB+ NVIDIA GPU. Smaller 7B-13B models run on more modest hardware but their output quality is meaningfully worse — fine for basic autocomplete, less good for substantive writing assistance. Most users do not have hardware capable of fully-local 70B-class models.

Local-model quality lags cloud frontier. Llama 3.2 70B is competent. It is also visibly behind GPT-5 and Claude Opus 4.7 on subjective writing quality, factual accuracy, and instruction-following. We ran the same prompts against local 70B and against the cloud APIs; the cloud outputs were noticeably better in roughly 70% of cases. For users whose primary concern is privacy, this is an acceptable trade. For users whose primary concern is output quality, the local-model path is not the answer.

Configuration is technical. Setting up Ollama, downloading models, configuring the plugin to point at the Ollama endpoint, tuning context-window sizes, and adjusting model parameters all require comfort with command-line tools and reading JSON configuration. We’re engineers and we found the setup straightforward; non-technical users will struggle. There is no out-of-box “just works” experience for the local path.

UI polish. The plugin UI is functional but not as polished as Notion AI’s hosted experience or ChatGPT’s inline editing. Settings panels are dense, the inline-completion visual indicators are basic, and some workflows require digging through nested menus. This is what a community-built open-source plugin looks like; it’s not a deal-breaker, but it’s a real difference from commercial AI products.

Comparison to alternatives in the category

Notion AI 3 is the polished hosted alternative. Better UI, better workspace-aware retrieval, weaker privacy proposition. For users not concerned about cloud-vendor data exposure, Notion AI is the more comfortable experience.

Cursor / VS Code with AI extensions are the developer-focused alternatives, but Obsidian’s note-taking workflow is different enough that they’re not direct substitutes. We mention them only because some users use Cursor as a writing tool.

Logseq with AI plugins is the closest direct competitor — a privacy-leaning markdown knowledge base with a community AI plugin ecosystem. Logseq’s AI plugins are less mature than Obsidian’s, and Logseq itself has had a bumpy commercial history that may give some users pause.

Apple Notes with Apple Intelligence is the cleanest Apple-ecosystem hosted option for note-taking AI, with strong privacy claims about on-device processing for many tasks. For Apple-ecosystem users with simpler note-taking needs, this is worth comparing.

Pricing

The plugin is free. Costs depend on which backend you use:

For light use (a few requests per day), cloud-API costs are typically a few dollars per month. For heavy use, costs can climb but you control the cap.

Verdict

Smart Compose is a competent AI plugin with one real differentiator: local-first model support that genuinely works. For Obsidian users who specifically value the privacy proposition — your notes never leave your machine, your AI runs entirely on your hardware — this is the recommendation. For users who’d prefer a polished hosted experience and don’t mind cloud-vendor exposure, Notion AI 3 or another commercial AI tool will deliver a better daily experience.

The plugin’s quality is OK, not exceptional. The local-model output is meaningfully behind cloud frontier models on quality. The setup is technical. The UI is functional rather than polished. None of these are surprising — this is what an open-source community plugin looks like — but they do mean that the product is for a specific user, not for everyone.

For that specific user — privacy-focused, technically comfortable, willing to trade output quality for data ownership — Smart Compose is the right answer in 2026. For most Obsidian users, the OpenAI or Anthropic API backend (which gives you cloud-frontier model quality with bring-your-own-key billing) is the more practical configuration.