SiYuan v3.7 CLI + Kernel Plugins + Block Drag: An Open-Source Note App’s Answer to the AI Era

If you use BibiGPT to turn Bilibili, YouTube, and podcast videos into structured notes with a single click, and you’re also a SiYuan (思源) user — the v3.7.x dev-channel releases are worth a close look. This isn’t a cosmetic refresh; it opens up “local AI automation” at the architecture level.

1. What Changed in SiYuan v3.7.x

In chronological order, SiYuan shipped the following capabilities in May 2026 across the v3.7.x dev series:

• v3.7.0-dev7 (2026-05-24) — Command-line CLI support (key milestone) / Kernel plugin system launches / New default theme and icon set
• v3.7.0-dev6 (2026-05-18) — Continued CLI iteration / Block editing enhancements (recursive collapse/expand for custom blocks) / Five new languages: Thai, Dutch, Indonesian, Hindi, Ukrainian
• v3.7.0-dev5 (2026-05-17) — Core plugin system / Block drag-and-drop (mobile + tablet) / Independent file-export solution for desktop and mobile (no browser dependency)
• v3.7.0-dev4 (2026-05-14) — Recursive block collapse/expand / IFrame Block performance improvements
• v3.7.0-dev3 (2026-05-12) — Mobile/tablet drag-to-reorder / Inbox drag images into editor / Multilingual expansions

Full changelog: SiYuan GitHub Releases

In one sentence: SiYuan v3.7 turns the note-taking app itself into a local Agent substrate drivable by scripts and plugins — without adding any built-in AI model.

2. What This Means for BibiGPT Users — My Take

下面这张手绘草图把刚才说的几个角色之间的关系一次画清：

配图：BibiGPT 团队为本文绘制（手绘风格）

At first glance this update seems unrelated to “AI notes” — there’s no built-in LLM button, no one-click summarize feature, nothing like Notion AI’s generation capability embedded right in a block. But zoom out one level and you’ll see SiYuan is taking a completely different route.

Take 1: CLI + Kernel Plugins = A Local Agent Programmable Layer

Notion AI and Tana voice-agent take the approach of embedding AI inside the product: you trigger it inside the product, consume results inside the product, compute runs in the cloud, your data goes to the cloud. That’s friendly for casual users, but the trade-off is that you have almost no control over “how the AI runs, whose servers it runs on, or which model it uses.”

SiYuan v3.7 takes the opposite approach: making the note-taking app itself a scriptable substrate. The CLI means you can create, edit, and search documents locally with commands like siyuan-cli --notebook="Work Notes" --create-doc "2026-05-28" — no GUI required. The kernel plugin system means plugins can hook into the document lifecycle inside the SiYuan process to do data processing, index building, and local vectorization — entirely on your machine, with whatever AI model you choose.

My take: This is a “meta-tool” strategy — SiYuan isn’t trying to pick an AI for you; it’s trying to become a local substrate into which you can plug any AI you want. Notion’s route is “AI serves you inside the product.” SiYuan’s route is “the product is infrastructure; you decide which AI, how it runs, and where your data lives.” For BibiGPT’s core user base — finance professionals, lawyers, doctors, and researchers — that’s a fundamentally different value proposition. I personally think SiYuan made the right call here. It chose a differentiation that Notion structurally cannot replicate: local-first + AI-programmable + zero data exfiltration.

BibiGPT’s power users and SiYuan’s power users overlap heavily. Both products attract Chinese-speaking power users who care deeply about data sovereignty: finance folks who can’t let investment research logic leak, lawyers and doctors who need client/patient information to stay local, researchers who don’t want unpublished ideas uploaded to the cloud. SiYuan v3.7’s upgrade directly lowers the technical bar for the “video comprehension → local notes → local AI re-processing” pipeline to a practically usable level.

Take 2: BibiGPT and SiYuan Are Upstream/Downstream, Not Competitors

When people hear “AI notes,” they instinctively put BibiGPT and locally-AI-capable note apps in the same competitive bucket. But look at where each product’s boundaries actually lie, and that judgment falls apart.

BibiGPT handles source-video comprehension: taking a 90-minute podcast, an online course, or a YouTube interview and using compute to turn it into structured text + chapter outlines + highlights + mind maps. That work depends on large-scale transcription, multilingual understanding, and video-timeline alignment — a compute-heavy task that runs in the cloud, and one SiYuan never intended to do itself.

SiYuan handles local knowledge consolidation and a programmable substrate: storing your notes in block structures, exposing an operations API via CLI and plugins, letting you decide how knowledge is organized and how AI processes it. That’s entirely local, with zero cloud dependency — and something BibiGPT never intended to do.

My take: SiYuan v3.7’s CLI + plugin system lets BibiGPT video summaries travel a completely local pipeline: BibiGPT does video comprehension in the cloud → pushes structured Markdown via webhook → SiYuan CLI writes it into the specified notebook locally → a kernel plugin calls a local model (e.g., Ollama) for cross-video RAG queries → everything after “video comprehension” is fully offline, zero data exfiltration. The privacy guarantee this pipeline provides is something Notion AI or any purely cloud-based note app cannot offer architecturally. You hand the video link to BibiGPT, BibiGPT does the understanding, and all subsequent knowledge processing happens on your own machine — that’s genuinely “saving your brain with compute while owning your knowledge.”

Mobile block drag-and-drop and recursive collapse/expand look like minor features, but for users who regularly organize notes on their phone, they’ve been a long-standing gap. On an iPad, dragging and reordering the chapter blocks that BibiGPT exports while watching a video is significantly smoother than before.

3. Hands-On Workflow: A Fully Local BibiGPT × SiYuan v3.7 Pipeline

The workflow below is designed around “video comprehension → local notes → local AI processing → zero privacy leakage,” covering the complete loop from video input to knowledge consolidation:

1. Paste a video link in BibiGPT, generate a structured summary

Open BibiGPT, paste a Bilibili, YouTube, or podcast link, and get a structured summary + chapter outline + highlights in 30 seconds. Output supports Markdown export — this is the source of the entire pipeline and the only step that requires cloud compute.

2. Export Markdown, push to the local SiYuan API via webhook

BibiGPT exports structured Markdown. You can use SiYuan’s built-in HTTP API (default localhost:6806) to receive the push:

curl -X POST http://localhost:6806/api/filetree/createDocWithMd \
  -H "Authorization: Token your-api-token" \
  -d '{"notebook":"Video Notes","path":"/2026-05-28-video-title","markdown":"# Title\n..."}'

This step is completely local — the network request goes from your machine to your machine, passing through no cloud service.

3. Use the SiYuan CLI to batch-write block structures into a specified notebook

The v3.7 CLI makes this step scriptable. You can write a simple shell script that watches BibiGPT’s export directory and automatically calls siyuan-cli to write new files into the appropriate notebook, auto-categorizing by video topic. For bulk organization this is more intuitive than raw API calls and easier to wire into automation tools.

4. Use kernel plugins + a local model for cross-video RAG queries

This is the capability the v3.7 kernel plugin system truly unlocks. You can use a community plugin (or write your own) to vectorize SiYuan documents, pair them with a locally-running Ollama or other local model, and query across videos semantically — for example, “Among these 30 finance videos, which ones discuss the relationship between interest-rate cycles and consumer recovery?” The semantic indexing and inference all run locally; original video transcripts are never uploaded to any server.

5. Manage chapter outlines with recursive block collapse/expand

BibiGPT exports usually contain multi-level chapter nesting — topic → sub-topic → key points → source quotes. SiYuan v3.7’s recursive collapse/expand for custom blocks lets you collapse an entire video outline with one click, viewing only the top-level topics and expanding only when you need the detail. Especially useful when organizing large topic-research projects spanning 10+ videos.

6. Mobile drag-and-drop: consolidate multiple video notes into one topic file

v3.7 dev5/dev6 added block drag-and-drop for mobile and tablet. When organizing video notes on an iPad, you can drag entire video-summary blocks to pull related content from different sources (Bilibili, YouTube, podcasts) into the same topic document — cross-video knowledge aggregation without switching to a desktop.

7. Fully local: video text never leaves your device

Only step 1 (video transcription + AI comprehension) requires cloud compute — and that’s precisely BibiGPT’s core value. Every knowledge-processing step from step 2 onward — storage, organization, secondary AI analysis, cross-video retrieval — runs entirely locally. For users handling sensitive content (investment-research recordings, meeting minutes, patient interview records), this pipeline’s privacy guarantee is something no purely cloud-based alternative can match architecturally.

Haven’t tried BibiGPT yet? try BibiGPT for free — paste a link and get a structured note in 30 seconds, with Markdown export that plugs directly into SiYuan’s API or CLI.

4. Video Comprehension + Local Knowledge Base + Zero Privacy Leakage

BibiGPT solves “turning videos into high-quality structured knowledge”; SiYuan v3.7 solves “letting that knowledge be re-processed by AI on your own machine.” Their boundaries are clean, and together they form a genuinely complete loop: video comprehension + local knowledge base + zero privacy leakage.

SiYuan didn’t take the popular route of “give users an AI button.” Instead, it chose to build the infrastructure that lets users plug in their own AI. That path is slower to walk, but once established, the degree of freedom and privacy it gives users is architecturally hard for other products to match. In a market where Notion AI and Obsidian AI plugins are everywhere, the “local-first + China-built open-source + programmable” combination actually becomes more distinctive precisely because everyone else is racing toward the cloud.

Want more note-app × BibiGPT workflow ideas? Check out our Note Apps blog category — every app has its own “1+1>2” workflow breakdown.

What kind of workflow are you building with SiYuan + BibiGPT? Drop a comment to share your workflow.