AI entered collaboration tools this week through privacy-first architectures and write-access agents, not just chat overlays.

The week in collaboration

The collaboration sector's AI integration wave landed differently this week than the usual chatbot-on-top-of-search pattern. The standout moves were structural: Readwise gave its AI write access to your entire reading library. Joplin shipped its first AI release with every feature off by default and local model support as a first-class option. Teable shipped a production-grade queued image generation pipeline across every app surface.

The common thread is tools designing AI integration with user control as the constraint, not an afterthought. That's a notable contrast to the broader market, where AI features ship enabled by default and privacy controls follow.

Leaders

Teable shipped queued AI image generation with model selection, progress tracking, and quota visibility across Chat, CLI, and App Builder — while also shipping Scheduled Routines and Composio integrations earlier in the week. The queued pipeline is production-grade infrastructure for a spreadsheet product: tracking, cancellation, and cross-surface consistency. The same week's Composio integration extends Teable's automation reach to hundreds of external services.

Readwise launched Global Ghostreader: an AI agent with write access to the entire reading library — tagging documents, taking notes, saving new content, and editing metadata — all governed by per-tool permission controls. Previous Ghostreader was scoped to the current document; Global makes it a full library operator. The governance model (per-tool enable/disable per-user) is worth noting: Readwise shipped library-wide write access with user-level controls rather than an opt-out.

Joplin 3.7 ships AI chat, semantic search, and MCP integration — its first AI-native release — built on a privacy architecture that disables all AI features by default, supports local models (Ollama, LM Studio), and requires explicit opt-in per feature. This is a meaningful design choice: building the on-ramp for privacy-sensitive users rather than assuming cloud AI as the default. Joplin's user base skews toward exactly those users.

Wiki.js 3.0 beta dropped AsciiDoc editing, blog pages, a native comments and discussion system, and a path-aware permission overhaul in a single release. The permission system is the most structurally significant: per-path overrides let admins grant editing rights to sections without exposing the full wiki. After years in beta, Wiki.js 3.0 is filling the gaps the 2.x series never closed.

Wildcards

Double introduced a two-tier AI model structure for its executive assistant product: a Standard model for routine tasks at 93% lower cost than the Expert tier, plus Auto routing that selects the appropriate model per message. The cost reduction is significant enough to suggest Double is repositioning AI from a premium tier into baseline product infrastructure — the same move happening in other AI-adjacent collaboration tools, but rarely at this scale of cost reduction in a single release.

SiYuan added direct Anthropic Messages API routing in v3.8.5-alpha.2, alongside an editable mind map view and database list view. Routing directly to Anthropic's API without a middleware layer is a deliberate privacy design — avoiding any telemetry or data retention from intermediary AI services. The combination of API-direct routing and local-first architecture positions SiYuan as the privacy-first end of the note-taking AI market.

Themes that compounded

• Privacy-first AI architecture emerged as a product design pattern: Joplin (all AI off by default, local model first) and SiYuan (direct API routing, local-first) both shipped AI features designed around explicit user control.
• Library-wide AI agents with write access: Readwise's Global Ghostreader is the clearest example this week — AI that acts across the entire knowledge base, not just the current document.
• AI cost tiering reached collaboration tooling: Double's 93%-cheaper Standard model with auto routing signals the commoditization pressure arriving in EA and assistant products.
• Feature-dense beta releases in knowledge tools: Wiki.js 3.0 and Teable's multi-feature sprint both packed structural improvements that had been deferred across prior cycles.

Watch this week

Watch Joplin 3.7's reception among its user base — the privacy-first architecture is the design bet, and user response will reveal whether privacy-by-default AI is actually what users want or just what they say they want. Also watch Readwise's Global Ghostreader adoption: library-wide write access for an AI is a different level of trust than document-scoped AI, and the per-tool controls will be tested by how users actually configure them in practice.