Distill app review — AI-native PKM with local-only limits (2026-05-19, 02:48)
Hands-on review of Distill, an AI-native PKM that just shipped: chat as the primary surface, chat messages as atomic units (vector-indexed, addable to lists), end-to-end encrypted sync, BYO model endpoint (small/medium/large). What kills it for this vault’s use case is the local-only stance — no cloud-runnable agents, no HTTP MCP, no mobile background scheduling — combined with a confusing four-tab UX (the “Inbox” is read-only and feels like a notification center, not an ingest funnel) and a planner that doesn’t belong inside a second brain. Trust/control-of-data also a concern even with markdown export. Verdict: won’t replace Obsidian, but several design ideas are worth stealing — atomic chat messages, lists-instead-of-folders, agents-in-sidebar, BYO-model endpoint.
[Speaker 1 - 0:01] So I just tried out this new app. It’s called Distill, and it’s basically an AI native personal knowledge management system that has this, well, a bunch of interesting ideas. First of all, like it’s kind of local, uh, and has a sync function similar to Obsidian. Then basically, the way you interact with it is that you just chat. So you have this chat window, which is the main, and then for each message that you send, you can also create a thread out of it. So usually would just work in these threads basically. And it seems like, I think that’s how it is, like every chat message also has like this little dot next to it, which means that it’s an entry in the personal knowledge management system, and you can add an entry to a list, for example. So well, basically, the atomic unit is a chat message in that, uh, software, and they are also indexed with, uh, semantic search, so vector index. And then you can, for example, add a— such a dot to a list. Well, agents can reason over them. So in Obsidian, you would have files as atomic unit, but here it’s these chat messages essentially, which is quite interesting. And then, yeah, I kinda find that pretty interesting, that concept. Then I find it interesting how they solve certain solu— certain things like, for example, encryption. So when you create your vault, you have to set a password, which is used for encryption end to end, I think, when they sync your data via their servers, which is a pro feature. And then they also have a built-in AI model basically that you get if you are subscribed. And you can also set your own model endpoint and s— choose like three models, like a small one, a medium one, and a big one. And that is also quite interesting. However, I feel like what’s annoying about it is that it does not have any cloud integration in the sense that you can let an agent run in the cloud. It is just local, right? And that kinda sucks because if you have scheduled task, tasks or something like that, you always have to keep your, uh, laptop open. And I’m not sure how these scheduled tasks would work on mobile either because you can’t schedule anything there. On iOS, for example, you can’t just wake up the app and just let it schedule stuff. That’s kind of bad, and th-that’s really not solving the issue that I have or would like to have solved. Then furthermore, the integration with other agents is also local only. So if you want to use Claude code with an MCP server to access your data, that also kinda doesn’t work because it is just local only. And if you are on the go and use Claude Code on the web, for example, it would require an HTTP MCP server, and that’s not possible with this. This is just some trade-off that the, um, the developers are making, I guess, from making it local first. But I’m not sure if that’s the right direction, honestly, because it gives off so much— well, so many good things about what could be done if it was cloud native. Furthermore, I noticed that it is quite… Well, I noticed this thing about my use case where I would not trust this software as much. So, uh, essentially, if I wa-were to dump my whole second brain in there, I would have to really trust the software that it’s, it’s not buggy, that it works correctly, that I can essentially trust it for the next five to ten years. And even though it is really polished and everything, I, I don’t know, it just feels weird to make the switch there because I’m not in control of my data anymore, I feel like. Even though there’s an export function, which can let you export your data to Markdown, it is still, yeah, feeling kinda weird in that sense, and that’s unfortunate. I’m not sure what’s the issue really here, but yeah. Anyways, um, other things that I do like about it are the visual styles. The, the design is pretty good. It’s on point. And, uh, I like the concept of chat messages being atomic units. I also like the concept of lists. So you can, instead of having folders or anything, you just have lists, which are essentially like you can s— put a message onto a list or a thread even, I think. And that makes sense, that kind of level of structuring stuff. You don’t need like folders or file structures. You just have lists and chat messages. So that is pretty good. Then also there’s… Well, what I find confusing is there’s this home view where you have like the chat, and then you have an inbox view, which is kinda weird ‘cause I think— like I thought of an inbox as something where I drop stuff that wa— that I want ingested into my vault or my brain. But this inbox is more like a Apple Notes thing where you see tasks due today and that kind of stuff. And you can’t actually write anything in there. It’s just, it’s read only, I feel like. And it’s… So I can’t create any new tasks manually there, and that feels really weird.I get why it’s done that way, but it just feels weird. So I would have thought that inbox is just throw stuff in here and the agent digests it, but it’s not what it is. And then there’s also the planner. So there’s actually four tabs. There is home, there is inbox, there’s planner. For some reason, when I click on inbox, it automatically takes me to the home view, but I have to click on inbox again to get to the inbox. So I feel it’s a bug currently, but while the inbox kind of has just four icons like new, read, snoozed, and dismissed. So it feels like a notification center, maybe like the one that I have in my vault. But yeah, it’s just weird if it’s called inbox, like I think of ingesting things. And then there’s the planner where you can actually write to-dos and even scheduled to-dos. But I don’t know. I’m not such big of a fan of having a to-do system in my second brain that feels like it should be another app that I can connect to my vault, but not inside of it. I know there’s many people who like that, but yeah, I’m not a big fan of it. Well, what I find interesting also is like you can mount folders. So you can select a folder on your drive where there’s markdown files in them, and then it will use that as extra context. It’s a pretty nice idea, I think. However, I’m not sure how I could import my Obsidian vault. So I’m just going to try that out now. Just going to drop in my whole Obsidian vault. One second. Where is it? Documents. Yeah, I’m not sure that’s working. So I just dropped it onto it, but it’s not doing anything. And I also cannot upload a folder to the Distill thing. Furthermore, I noticed that after signing into my vault, there’s nothing getting synced. So maybe I have to subscribe to the sync plan, but still I had to log in with my vault password, but it’s not syncing anything. And that’s just really weird. So yeah, I like the promise and the overall design decisions where it’s going, but there’s just, I don’t know, so much. There’s just many weird design decisions that don’t fit how I would like it to work. Then also it’s not extendable in the sense that you can add any extensions or stuff to it. So that is also really annoying. So yeah, I guess the bar for such a software is really high and I’m not even sure if the things that I want to get out of it are compatible with what’s actually possible. But I think it’s really interesting in some, in the design sense. Like I like the design. I like how agents are present at the sidebar. I like the chat messages being atomic units, but then there’s other weird stuff that just doesn’t make any sense or I just find annoying. So I don’t think I can switch from Obsidian to this new tool, which looks pretty cool actually.
Design ideas worth stealing
- Chat messages as atomic units, vector-indexed. The unit Obsidian/vault PKM treats as a file (or block) is the unit Distill treats as a chat message. Each message is searchable, addable to a list, reasonable-over by agents. Reframes the database: it’s not a graph of notes, it’s a stream of utterances with structure layered on top.
- Lists instead of folders. No file structure, no PARA hierarchy — just lists, and threads or messages get added to lists. The structuring primitive is set membership, not path.
- BYO model endpoint with size tiers (small / medium / large). Subscriber gets a built-in model; advanced users plug their own. Three named sizes is a nice abstraction over “pick a specific model name” — useful UX pattern for a multi-provider, model-agnostic setup.
- Agents present in the sidebar. First-class presence at the UI surface, not hidden behind a settings panel.
- End-to-end encryption with a vault password. Sync via their servers, but they can’t read your data. Aligns with the “trust the software for 5–10 years” worry — pragmatic compromise.
The specific dealbreakers
| Decision | Why it kills the use case |
|---|---|
| Local-only agents | No cloud-runnable routines. Laptop must stay open for scheduled tasks. iOS can’t wake the app for scheduling either. Same gap Open-Source Model-Agnostic AI Platform and OpenClaw for PKM both call out at the cloud routines layer. |
| Local-only MCP / agent integration | No HTTP MCP server, so Claude Code on the web can’t talk to it. The skills-as-public-API surface argued for in 2026-05-10-Inbox-style PKM app - input funnel and iOS sync challenge is structurally impossible here. |
| Trust / data control | Polished, but locking five-to-ten years of second brain into a young opaque app feels weird even with markdown export. Same loss-of-control instinct that keeps the current setup pinned to Obsidian + git. |
| Inbox is read-only | The “Inbox” tab is a notification center (new/read/snoozed/dismissed), not an ingest funnel. The mental model collision is brutal — this vault’s whole point is ingest funnel. See process-inbox and 2026-05-10-Inbox-style PKM app - input funnel and iOS sync challenge for the opposite stance. |
| Planner inside the second brain | The to-do system feels out of place; should be a separate app connected via API, not bundled. |
| No extensibility | No plugins, no extensions. Closed surface. |
| Sync not working | Logged in, vault password accepted, no actual sync — possibly subscription-gated, possibly a bug. Unclear from the UI. |
| Folder mount didn’t import vault | Dropped Obsidian vault on the app, nothing happened; no folder upload affordance either. |
How this lines up against the OpenClaw for PKM track
Distill is the shipped, opinionated, local-only version of a chunk of what this track is sketching. Mapping back to the six challenges from 2026-05-06-Standalone PKM app - cons of Obsidian setup, sync, UI focus:
- #2 Sync — Distill chose E2E-encrypted server sync with a vault password. Concrete pattern worth knowing exists, but doesn’t address the agent in the cloud gap.
- #3 Local-first markdown vs. databases — Distill went database (chat messages are the unit, markdown export is escape-hatch only). Opposite of the markdown-as-format pivot in 2026-05-10-Inbox-style PKM app - input funnel and iOS sync challenge.
- #4 UI — Distill’s “chat as main, lists as structure” is one answer; the funnel-first answer from 5/10 is another. Distill puts chat first; the OpenClaw direction puts ingest first.
- #5 Core loop — Distill’s loop is chat → message-as-entry → list → agent-reasons-over. No equivalent of
process-inboxstyle routing into a structured graph. - #6 Cloud routines — Explicitly absent in Distill. The most expensive missing piece, and the one the OpenClaw direction treats as table stakes.
Open questions raised
- What does an HTTP MCP server on top of a PKM look like end-to-end? Distill’s local-only MCP gap is a concrete proof that “skills exposed to outside agents” is non-trivial to ship — what would Kumo’s MCP surface need to add to fill this gap for a PKM tier above it?
- Is the atomic unit = chat message abstraction a real upgrade over atomic unit = note/block, or does it just shift the file-organization problem one level up (now you organize lists instead of folders)? Worth a follow-up note.
- Is “agents in the sidebar” the right UI primitive, or does it become noise once you have many?
- What’s the right trust story for a PKM that wants to hold 5–10 years of someone’s second brain? Markdown export is a floor, not a ceiling — what else?
Related
- OpenClaw for PKM — parent concept page; this memo is a comparable-product datapoint. Distill is roughly the local-only, chat-first slice of what OpenClaw wants to be, minus the proactive background agent and cloud routines.
- AI Native PKMs — predecessor idea (chat + graph DB); Distill is one shipped instance of that pattern, with chat-messages-as-atomic-units replacing the graph DB ontology question.
- 2026-05-10-Inbox-style PKM app - input funnel and iOS sync challenge — sister memo where the inbox funnel and skills exposed to outside agents shapes were sketched; Distill’s read-only Inbox tab and local-only MCP land on the opposite of both choices.
- 2026-05-06-Standalone PKM app - cons of Obsidian setup, sync, UI focus — six numbered challenges this memo grades Distill against (#2 sync, #3 local-first markdown vs. DB, #4 UI, #5 core loop, #6 cloud routines).
- 2026-05-06-Standalone PKM app - yjs CRDT sync stack — sync stack the OpenClaw track is considering; Distill’s E2E-with-vault-password is a different point in the same design space.
- 2026-05-06-Standalone PKM app - Cursor SDK as harness layer — harness question; Distill answers it with a built-in model + BYO endpoint instead.
- 2026-05-10-Voice note processing - Monologue MCP integration plan — stay-inside-Obsidian answer to the same problem space; Distill is the leave Obsidian answer, and reads here as a vote against leaving.
- My AI native Obsidian Setup — the current setup this evaluates against; “I don’t think I can switch from Obsidian to this new tool” is the conclusion, and this page is what wins.
- Open-Source Model-Agnostic AI Platform — the cloud routines + BYO model platform framing; Distill ships the BYO-model piece but not the cloud-routines piece, illustrating the gap that page names.
- Agent-Native Software Architecture — the architectural pattern; “agents in the sidebar” + “lists as the structuring primitive” are two surface choices in that pattern’s design space.
- Kumo — adjacent storage tier; a PKM like Distill could in principle sit on top of a Kumo-style cloud filesystem. The HTTP MCP gap Distill leaves open is exactly the surface Kumo provides.
- Posts — possible venue if this gets written up as a public review.