Today we’re releasing Linkly AI v0.4.0.
This is the biggest update in Linkly AI’s two months since launch — not just because it ships a lot of new things, but because it’s the first release where Linkly stands on its own. You no longer need to plug Linkly into Claude, Cursor, or OpenClaw to “have a conversation.” Linkly AI now has its own Chat, which marks a real evolution in our product positioning.
Open it up, and you can directly talk to — and do research on — the thousands or tens of thousands of notes, PDFs, journals, and project records sitting on your hard drive.
Why We Built Yet Another Chat
For the past few releases, we’ve deliberately put our energy into the tooling layer — MCP tools, CLI, Skills — letting Agents like Claude Code, Cursor, and OpenClaw seamlessly search, read, and integrate your local files.
Our judgment back then was: users probably don’t need yet another Chatbot — they need better tools.
After more than 20 iterations, our tooling layer has matured. But after some time in beta, we noticed two issues that deserved more attention:
- A large number of users aren’t using Claude Code, Cursor, or any other Agent at all. They install Linkly AI, finish indexing tens of thousands of documents, and then realize there’s no simple entry point to actually leverage that knowledge.
- Those Agents still depend on cloud models, which can’t satisfy some users’ extreme privacy requirements. They need a knowledge assistant that runs entirely locally.
A pure tooling-layer positioning was holding back too many users. We needed to provide a more direct entry point and let users decide for themselves whether to start there, or integrate Linkly into their own Agent. So adding a Chat became the natural next step.
What Makes Linkly AI Chat Different
We didn’t want to ship yet another lookalike Chat. Linkly AI Chat goes deep on one thing only: conversations and research grounded in your local files. Compared to general-purpose Chatbots like ChatGPT and Claude, here are the three things that genuinely set it apart.
1. Seamlessly Integrated with Your Local Files
This is the most fundamental — and most underestimated — difference: ChatGPT and Claude can never reach into your hard drive.
Every time you want them to read your stuff, you have to manually copy, paste, or upload, then worry about how they’ll handle it afterwards. Linkly AI Chat just works — the moment you open Chat, the thousands or tens of thousands of notes, PDFs, journals, and project records on your drive are already its context. And all of that data stays local. Nothing gets uploaded, no data security headaches.
The first time I used it, I asked it some questions about myself. It immediately surfaced a note I’d written more than ten years ago and reconstructed the whole context for me. I’d never imagined those memories would resurface in this form — it was the first time I felt the magic of local-first.
2. Powerful, Accurate Retrieval and Cross-Document Synthesis
“Searching your local files” isn’t novel — what’s hard is searching accurately, deeply, and trustworthily. Chat doesn’t use the usual “chunk + vector database” recipe — we walked away from traditional RAG a long time ago and replaced it with a stack we built in-house, designed for Agents to explore local files on their own:
- Outline Index: Each document gets a structured “outline” first, and the AI flips through chapter by chapter — context stays whole, and Token usage stays lean.
- Agentic RAG: The model itself decides what to search for, when to keep digging, and when to stop. It’s not a fixed workflow like the one inside Perplexity, which is exactly what lets it pull off multi-round, Deep Research–style retrieval and analysis.
- Sources on every answer: Every response automatically lists the specific files it cited, with full source-passage viewing coming in a later release. For a private corpus this matters enormously — you have to know which file the AI’s answer came from, otherwise it’s just “another AI that makes things up.”
- Built-in Skills: Linkly AI ships with a carefully tuned set of tools, but every model uses tools differently. We’ve baked the right tool-combination patterns into built-in Skills based on extensive practice — so you get consistent performance even when you swap to a different model.
I often test it with one specific scenario: ask it to write me a résumé. It’s a textbook cross-document synthesis problem — there’s a clear timeline, and every fact is easy to verify. My computer holds résumé drafts, self-introductions, project retrospectives, and promotion materials from different periods, different roles, different contexts. It pulls all those scattered pieces of “you” together and stitches them into a more complete version than I could’ve done by hand. ChatGPT can’t do this — it simply can’t read those files.
3. Multi-Provider Support, Truly Offline-Capable
Linkly AI Chat ships with the Linkly Official Curated Model Service, ready to use with zero configuration. Two trial models launch with it: Qwen 3.5 Flash and DeepSeek V4 Flash, both chosen for their balance between tool-calling quality and cost.
If you’d rather use your own model service — OpenAI, Anthropic, DeepSeek, Ollama, LM Studio — they’re all freely configurable. Unlike ChatGPT or Claude, you’re not locked into a single model.
Going further: if you connect to a local model via Ollama or LM Studio, you can run Chat 100% offline — the entire conversation never leaves your computer. That’s invaluable for users and scenarios with extreme privacy requirements: lawyers, doctors, journalists, in-house researchers at enterprises.
What It Doesn’t Do Well Yet
We have to be honest: Linkly AI Chat is still pretty bare-bones. Outside of “conversations and research grounded in local files,” it falls short of mature Chatbots in pretty much every dimension:
- General Q&A — the knowledge breadth doesn’t match ChatGPT or Claude
- Code generation — far behind Claude / Codex / etc.
- It also doesn’t have internet browsing yet
- It also lacks context window compression, MCP integration, third-party Skills, and other standard features
For now it only suits the vertical scenario of “Q&A on your local private corpus” — but we’ll keep iterating on it quickly. If your use case is more complex and you want a more powerful model, we’d suggest using a more mature Chatbot, or treating Linkly AI as a tool and integrating it into them.
New: Data Privacy Panel — No More Black Box
In our beta, many privacy-sensitive users were worried about data security. As a local-first AI app, we know data security is our lifeline, and fully offline operation is the long-term goal. But local compute is limited for many users — for a small set of features, the local experience just isn’t good enough yet, and we have to lean on the cloud. What we can do, however, is make all of that fully transparent.
So in v0.4.0 we added the Data Privacy panel, which centralizes and transparently shows where your local data flows, visualizing every way Linkly AI handles your data.
We also added a small but flexible toggle: show full file paths.
- MCP tool results and CLI output hide full paths by default — to avoid leaking your directory structure, username, or project codenames to cloud models
- When you’re working in your own local environment and need an Agent to operate on files directly, you can turn it on yourself
Conservative by default, opt-in for openness — that’s our consistent stance on privacy.
Other Improvements Worth Mentioning
- In-app Changelog view: Read the latest version updates right inside the launcher, no browser needed.
- Local model download progress: Local embedding / model downloads used to leave you staring at a frozen spinner. The AI settings panel now shows clear download progress and status.
- New Pin button on the Launcher: The launcher used to disappear the moment it lost focus, which broke a small but real “look-up-while-doing” workflow. With Pin enabled, the launcher stays put until you close it yourself.
- Three new languages: Japanese, French, and Spanish.
- MDX file support:
.mdxfiles, common in blogs and technical documentation, are now indexed and searchable. - Linkly Official Curated Model Service launches — new users get going out of the box.
Bug Fixes
- Fixed an issue where the embedding model failed to download in certain network environments, blocking initial indexing
- Fixed a settings page scrollbar overlapping with content
- Fixed item selection in the launcher occasionally drifting
- Fixed several crashes and hangs on the Windows client
- Fixed one-click Claude Desktop integration that previously failed; the app now guides you through installing Linkly CLI on first integration if it’s missing
Full changelog is available here.
If you haven’t tried Linkly AI yet, you can download it for free from the official website — available on macOS, Windows, and Linux.