Repository navigation
Where Tine is going #645
Replies: 10 comments 27 replies
|
so consumers of a static export can run arbitrary, ad-hoc queries? this suggests a query-engine that can work in-browser, probably by compiling the existing rust query engine to WASM. interesting!
I don't, due to privacy concerns with cloud vendors. i've done some experiments with local open weight models but they are weak.
I have no immediate use for this, but you should build what you need :) @EllisMorrow: Note he said syncing via git, not github. You could use a git-remote-gcrypt intermediary remote (seems a bit clunky) or p2p sync (then devices need to be both online at the same time to sync) |
This is amazing, I was already using Tine a lot like this, by having it edit my md and it wasn't so bad.
Can absolutely validate this pain-point. After using Calude to make my notes for a while, my notes started to feel like AI slop themselves. And that was a terrible feeling. Since then, my workaround is that I ask Claude to always add #AI-note to the top level bullet on where it makes the edits. My plan was to use Custom CSS to style is differently, but the However, here is the challenge I've been contemplating about. Consider this note: What do we do? 🤔 Now the human text is credited to Claude. My thought: Explained:
The concept:
The mechanisms
What are you thoughts? |
This would be great, hopefully won't get messy in code. As an oracle, GitHub Copilot Agent in VS Code had an absolutely splendid way to track and approve/deny agent changed lines. And afaik they did it without interfering with git. Maybe it could be useful if the Copilot integration code is open-to-read. Do check with Claude on this, I may be off with my pointer. |
I use it a lot. I think you nailed my pain points with Authorship and Approval. Whats left Claude tends to be lazy at atomically inserting at multiple points and intead append all-in-one blobs Suppose I ask "Can you add the Good things about each fruit?" Agents tent do do Instead of This is a caricatured example, if you test it, it will probably split it well, but when it involes changing a lot of separate things, it becomes lazy Why? I am not sure. Could be bad training. I managed to indoctrinate my agents with some harsh prompting to do like I desire, but its not perfect. A theory: the cost of multiple batched changes Changing many (such as 10) is complicated. Usually harnesses can do multiple calls at once without rethingking, but if the previous call shifts line numbers around, things get messy, A single miscalculation in agents 'head' and subsequent edits fail. So basically, whatever API/CLI/MCP/SDK Tine provides for agents, it would be great if it enabled (somehow) to easily do sequenced edits on the same md with a single batch job. This is easy if each Block has an ID, but its not the case in LogSeq, each item gets an ID only when and ID is needed. |
|
Regarding the first section on agents: If so, would it be possible to make the MCP server a thing which you have to enable first? I'm asking because I personally wouldn't want any agents/AI near my graphs, and in that case there is no need to run an MCP server in the background. |
|
2: I prefer using queries to define what to share (with no visual component), but queries that can include both blocks and pages. For example, I might want to share an entire page, but also specific blocks from another page, while excluding certain blocks from a page. How could this work? We could allow users to define multiple queries specifying what to share, in addition to adding individual pages and blocks. If a block from a page is selected, its entire hierarchy is exported (parent blocks and the page itself as the root block). Then, for each query or page, we could run an extraction query to remove pages or blocks from the initially selected pages and blocks. I could develop this idea further if it sparks any interest. 3: I'd really like to see this implemented, as it would solve many limitations and could be used to synchronise different devices. It would be even better if we could choose the synchronisation policy, for example, automatic synchronisation for one repository and on-demand synchronisation for another. It could even be used as a collaborative tool, tracking who authored or modified each block or page, and so on. All of this could be achieved without cluttering the Markdown files, by extracting this information from the repository. Commits could also have semantic messages in the commit log, such as "renamed X to Y" or "added property to this page or block", and so on. 4: I'd approach this one carefully. Tine is evolving very quickly, and I think we should establish a solid foundation before opening this door, because it could lead to technical debt very quickly. I've never been a fan of plugins. I understand why other people need them, but using plugins to fix or work around problems in an application's design (Obsidian is the best example that comes to mind) isn't the solution. How many users have said something along the lines of: "I like Logseq because it works without plugins, and its workflow fits the way I work..."? |
|
Regarding “4. Plugins,” Obsidian is arguably one of the strongest note-taking applications when it comes to plugins, and there are many valuable lessons to learn from its approach to plugin development. I'd like to share an interview with Obsidian CEO Steph Ango on the Decoder podcast, where he discusses the role of the plugin system in Obsidian's development: [Obsidian CEO Steph Ango interview — The Verge](https://www.theverge.com/decoder-podcast-with-nilay-patel/760522/obsidian-ceo-steph-ango-kepano-productivity-software-notes-app?utm_source=chatgpt.com) He makes several important points:
One especially valuable approach is a three-stage development model:
|


Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Written by Claude (Opus 5.5) from Martin's notes and decisions; read and approved by Martin.
In #422 Martin said that once the direction was clear enough he would write it down visibly. This is that post: I wrote it from his vision notes and his answers to my questions. Nothing in it is a schedule. Right now the work is making Beta solid enough to become 0.7.0 (see #617); stable gets only severe fixes until then. This is a map of what comes after, and what feature requests are measured against.
The short version: Tine is a workbench where you and your agents work on a Markdown graph that stays yours. Any slice of it can be handed to anyone as a working copy of the app.
Most of the individual pieces exist somewhere already, and I'll say where. Martin thinks the combination is what's new, and it is what he wants for his own work: research with agents, a large literature atlas he is building for a project, and teaching.
1. You and your agents, on the same page
What Tine will not do: ship a chatbot, bundle a model, or grow an AI sidebar. You bring the agent you trust (Claude Code, Codex, a local model, whatever comes next) and run it next to Tine. Tine gives agents a command line first. Its commands ask Tine's own query engine instead of grepping your files, and they make edits correctly (block ids, links, embeds) through the same guarded save path the app uses, so even a weaker model can edit your graph without learning Tine's syntax. For agent apps without a shell, an MCP server built from the same commands follows, and it can also tell the agent what you're looking at.
What matters is what happens after the agent writes. Martin's own flow is not "accept or reject". He lets the agent write, reads what it did, and pushes back where he disagrees. Two things are missing for that today:
author:: claude, which Tine renders as clearly not yours. It is plain Markdown, so other tools see it too, and in a team graph shared over git it attributes humans just the same.An open question: margin or sidebar? A cheaper first version would keep each thread as a collapsed child of its block, show a badge, and open the thread in the right sidebar Tine already has. That works with split panes and on phones, but you don't see comments as you scroll, which the margin gives you. Martin leans toward the margin; we'd like to hear which you'd use.
If an agent run went wrong as a whole, Tine should show what that run touched and revert it in one step. This is for graphs kept in git: each command-line run becomes a commit with the agent as author, so the history already holds what changed. Tine reverts block by block with the same merge (Concord) that handles sync, so anything you edited after the run is kept and shown to you, not overwritten.
The nearest thing I know of is Logseq's new database version, where agents can answer in comment blocks. The plan differs in two ways: everything stays plain Markdown in your files, and the comment is anchored to the passage it answers. Google Docs has the margin and many tools have MCP; the combination is the point.
2. Exports that are the app, and sharing without a server
Tine already exports static HTML, and since the query work you can export a single query's pages as a standalone site. The next step is an export that is the read-only app: Ctrl-K, hover previews, backlinks, and queries a reader can type themselves, running in their browser. To be fair to prior art: Logseq's "Export public pages" and TiddlyWiki have had live queries in exports for years. What's rare is a reader asking new questions of a static export, and that is what Martin wants for publishing research.
Choosing what to share. A selection is just a query, with one addition: "these pages plus everything within k links". You'll be able to pick it in the graph view and Tine writes the query for you. We haven't designed that view yet: the sketch below (search and seed pages at the top, a lasso, hop and filter controls at the bottom) is only a starting point, and ideas are very welcome. Graph tools converged on this "seeds plus k hops plus filters" model long ago; nobody seems to use it to decide what gets published. A query can be marked as exported, and one button or one CLI command re-exports all of them, so a nightly job keeps your published work current. Before re-exporting, Tine tells you which pages joined or left the selection, because a selection defined by links grows by itself.
Sharing without a server. One button packs the selection, puts it in a secret gist on your GitHub, and gives you a link with the decryption key in the part of the URL that never reaches any server. Tine runs no hosting and stores nothing. Each piece has shipped somewhere (SoloMD, Datasette Lite, Share Note); a multi-page working app shared this way, as far as I can find, has not.
Mounting. The same pack can be mounted read-only into another Tine: Ctrl-K and
[[find its pages next to yours, and you annotate it in your own graph. One thing in it I haven't found anywhere: backlinks from the mounted corpus into your own notes ("who in this atlas cites the concept I'm writing about"). It's designed to stay fast at about a hundred thousand pages without loading them into memory. What Tine is not building is "follow other people's graphs"; that idea has been rebuilt for fifteen years with little uptake, while mounting reference libraries and pulling sources into your own notes is what people actually use.@almmos, on #422: there will still be no access-control system inside Tine. Reading is by possession of a link or a file. Live shared spaces with membership and keys are off the map until someone has a concrete need that sharing via GitHub doesn't cover.
3. Syncing through git
Tine doesn't run a sync service. Some of you already keep your graph in a git repository, and the plan is for Tine to sync it for you, the way SoloMD's git autosync does (and the git-enabled Logseq fork nataloko wrote years ago). It's opt-in, only for graphs that are already repositories, and it runs on phones too: git is built into the app, so nothing needs to be installed, and you sign in once with a token. On a phone it syncs when you open or leave the app, since phones don't allow syncing in the background.
The first version only ever fast-forwards. If two devices have diverged, it stops and says so, and your local copy stays exactly as it was. The second version merges a diverged history page by page with the same block-level merge (Concord) that already handles Syncthing and outside edits. git helps here: it knows the exact common ancestor, which Syncthing has to guess. Blocks that genuinely conflict keep both copies and wait for you. Paired with that comes a small "changed since you last looked" highlight per page, so you can see what another device, a collaborator or an agent changed.
4. Plugins that agents write
Martin expects most plugins to be written by agents for one person's needs, the way Claude Code users ask their agent for a mod and get it hot-loaded. So the plugin API will grow an explicitly experimental, typed tier: page views, sidebar panes, render hooks and query results, built and reloaded locally with the tooling an agent needs. Experimental plugins stay on your machine; stable ones go to the registry. Plugins draw by returning data (a tree of views), never by touching Tine's internals. That way a view survives into an export without running anyone's code in a reader's browser. One rule is absolute: no code ever arrives with content. Opening a shared link or mounting a pack never runs a plugin. Two incidents this year, malicious plugins spread through shared Obsidian vaults and a wiki import that executed code, are exactly that failure.
What this means for feature requests
Most requests are for parity with other tools. They're welcome and some will happen, but they come after this map, because they neither make Tine different nor change Martin's own work much. Requests that make Tine a better place to work alongside agents, or to publish and reuse what you've built, will find him receptive. "Add an assistant" will get the answer "bring an agent; Tine will talk to it." And none of this is scheduled: the work happens when Martin points at it, possibly in a different order than written here.
What comes next, concretely: the order of work, issue by issue, is in the companion roadmap post. That's where to argue about priorities; this post is about direction. How Tine is built, and what we're trying there, is a separate post: How Tine is built.
Martin would like to hear from you. If you already point an agent at your notes, what would you want it to see, and how do you want to answer it? Would you mount a corpus someone else publishes, or publish one yourself? What would you share, and with whom?
All reactions