I have seven inboxes. Maybe you do too.
YouTube Watch Later, accumulating since 2022. Work emails, trapped in a corporate Outlook prison. Tasks scattered across daily notes, GitHub issues, and two different AI agent sessions I started and forgot about. A Jira backlog that works for work but connects to nothing else. And the worst inbox of all: my own head, holding tasks for months because there was nowhere fast enough to put them.
The Getting Things Done methodology starts with one unbreakable rule: capture everything. If it’s in your head, get it out. If it arrives anywhere, route it to one place. A universal inbox. Zero friction. Trust the system.
So I asked myself a simple question. Has nobody solved this? In 2026, with self-hosted everything and AI in every CLI, is there really no single tool that captures tasks, links, emails, files, and voice notes into one inbox?
I searched GitHub, Reddit, YouTube, and AlternativeTo. Here is what the landscape actually looks like.
The dream: one tool, every modality
The ideal capture tool handles six modalities. Not five. Six.
| Modality | What it means | Why it matters |
|---|---|---|
| Tasks | Quick action items, todos, reminders | The most frequent capture type. If you can’t task-capture in under 5 seconds, you won’t do it. |
| Links | URLs to read, watch, or reference later | The second most common. YouTube videos, articles, GitHub repos, tweets. |
| Notes | Fleeting thoughts, ideas, observations | Unstructured. Needs zero fields, zero friction. |
| Files | Screenshots, PDFs, images, documents | The “I’ll need this later” capture. Phone screenshots especially. |
| Forwarded messages, newsletters, receipts | Most people’s highest-volume inbound channel. | |
| Voice | Spoken notes, transcribed thoughts | Essential for driving, walking, shower-thoughts. |
A single tool that does all six. Self-hosted. Privacy-respecting. Open source. That was the search.
What’s actually out there
I evaluated every candidate against four criteria: how many modalities it covers, whether it’s truly self-hosted, whether it stores data as plain markdown files (portable forever), and how fast capture actually is in practice.
| Tool | Tasks | Links | Notes | Files | Self-hosted | Markdown files | |
|---|---|---|---|---|---|---|---|
| Tandem GTD | ✅✅ | ❌ | ✅ | ❌ | ✅ | ✅ Server | ❌ (Postgres) |
| DropMind | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ Server | ❌ (SQLite) |
| Trove | ❌ | ✅✅ | ❌ | ❌ | ❌ | ✅ Server | ❌ (SQLite) |
| tududi | ✅✅ | ❌ | ✅ | ❌ | ❌ | ✅ Server | ❌ (SQLite) |
| MindVault | Weak | ✅ | ✅ | ✅ | ❌ | ✅ Server | ❌ (SQLite) |
| Karakeep | ❌ | ✅✅ | ✅ | ✅ | ❌ | ✅ Server | ❌ (SQLite) |
| NeumanOS | ✅ | ✅ | ✅ | ❌ | ❌ | ⚠️ Browser | ❌ (IndexedDB) |
| nb CLI | ✅ | ✅ | ✅ | ❌ | ❌ | N/A (local) | ✅✅ |
| Syncthing + Obsidian | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ P2P | ✅✅ |
The pattern is unmistakable. Every tool that handles multiple modalities uses a database. Every tool that uses plain markdown files handles fewer modalities. Nobody has reconciled the two.
But there are three tools worth examining closely. Each gets something right that the others miss.
Tandem GTD: the methodology purist
Tandem is the most complete GTD implementation I have ever seen in open source. It does not just have an inbox. It has automatic next-action cascading, context views with energy and time filters, all six Horizons of Focus, a guided weekly review with an optional AI coach, and even a personal wiki with backlinks.
The quick capture is genuinely fast. A global hotkey opens a capture bar. Type. Press Enter. The item lands in your inbox for processing later. It supports natural language date parsing. It has an MCP server so Claude or any compatible AI can read and write your tasks.
Tandem also has email capture. A Cloudflare Worker receives inbound email and routes it into your inbox. That is one of the hardest capture sources to solve, and Tandem solved it.
But. Tandem stores everything in Postgres. If the project dies, your data is a database dump, not a folder of readable markdown files. You are locked into the application. For a tool that organizes your entire life, that is a non-trivial bet.
DropMind: frictionless, temporary, perfect for what it is
DropMind does exactly one thing. It captures. Links, notes, screenshots, files. From iOS Shortcuts, Android Share, a browser bookmarklet. No folders. No tagging. No decisions. Just drop it and move on.
The philosophy is deliberate. DropMind is explicitly not a note-taking system. It is a temporary capture space. You capture on your phone while commuting. You process on your laptop later. Keep it, move it elsewhere, or delete it.
This is the closest I found to a true universal capture front-end. But it has no task management. No recurring items. No processing pipeline. It is a brilliant first step in a GTD workflow that stops at step one.
Trove: the link router with taste
Trove is the newest of the three, built in early 2026 by Jayden Kim. It is a self-hosted link inbox. But calling it a bookmark manager undersells it.
Trove has a triage mode. You sit down, press T, and it feeds you links one at a time. Keyboard shortcuts send each one to its destination. 1 sends to Readwise Reader. 2 creates a Things task. 3 routes to Obsidian. You can blaze through 50 links in three minutes. Nothing gets stuck.
The plugin system is where Trove shines. You define JSON manifests that describe where links should go and what HTTP request sends them there. Built-in plugins ship for Reader, Things, Notion, Obsidian, and Apple Reminders. A template engine lets you transform link data into the exact format each destination expects.
Trove also has an MCP server. Your AI agent can search your link archive and save new links through natural conversation. For the Claude Code users among us, that is genuinely useful.
But Trove is links only. No tasks. No files. No voice. It solves one modality with exceptional polish and ignores the other five.
Why markdown files matter
Every tool above stores data in a database. SQLite. Postgres. IndexedDB. That works. Until it does not.
The single most important property of a capture system is that your data outlives the tool. I have migrated note systems four times in ten years. Evernote to Apple Notes to Notion to Obsidian. Each migration cost days of exporting, reformatting, and link-repairing. Plain markdown files on disk have never required a migration. They are readable by any text editor on any operating system, now and in 2056.
This matters especially for a GTD system. Your trusted system contains your projects, your waiting-for list, your someday-maybe ideas, your reference material. If the tool dies, the system dies with it. And rebuilding a trusted system from scratch is months of work.
The tradeoff is real. Databases enable features that are genuinely hard in flat files: full-text search, relational queries, collaborative editing, conflict resolution. Markdown files give you permanence and portability but limit what the tool can do without a database layer on top.
No. Not yet. But the pieces are closer than they were two years ago. Tandem gets GTD right but uses Postgres. DropMind gets frictionless capture right but stops at capture. Trove gets link routing right but only does links. nb gets the CLI and markdown approach right but has no mobile or web capture.
Here is what I ended up building for myself. It is not a single tool. It is four layers.
| Layer | Tool | What it does |
|---|---|---|
| Sync | Syncthing | P2P file sync across laptop, phone, and homelab VPS. No cloud. |
| Capture | Shell function + HTTP server | capture "buy milk" appends to today’s markdown inbox. Takes under two seconds. Phone shortcuts POST to the same endpoint. |
| Processing | pi agent + cron | Morning review processes overnight captures. Evening review compares planned vs done. Weekly synthesis scans for stalled projects. |
| Editing | Obsidian + Markor | Full-featured PKM on laptop. Minimal markdown editor on phone. Same files. |
Is it elegant? No. A single person’s shell script and a 30-line Python HTTP server are not the universal capture tool the world needs.
Does it work? Yes. Capture is under two seconds from terminal. Phone capture works through a Shortcut. Syncthing ensures my vault is identical on every device. The AI handles the drudge work of classification and flagging while I make the decisions. And every single byte lives in plain markdown files that will outlast every tool in the stack.
What this search taught me
The universal inbox problem is harder than it looks. Not technically. Technically, a database table with a type column solves it in an afternoon. The hard part is that different capture modalities have fundamentally different requirements.
A task needs to be captured in under two seconds or you will not do it. A link needs automatic metadata extraction or you will not find it later. A file needs thumbnail generation. An email needs parsing and deduplication. Voice needs transcription. Each modality demands a different capture interface, a different processing pipeline, and a different storage schema.
The tools that try to do everything end up mediocre at everything. The tools that do one thing well are genuinely excellent but leave you stitching five of them together. This is not a failure of engineering. It is a reflection of the problem’s actual shape.
The second thing this search taught me: capture speed is the only metric that matters. You can have the most beautiful inbox in the world. If it takes more than five seconds to capture a fleeting thought, you will not capture it. You will tell yourself you will remember. You will not remember. The thought will die in the gap between intention and action.
This is why I landed on a shell function. It is ugly. It has no UI. It cannot extract link metadata or generate thumbnails. But it takes 1.2 seconds from thought to captured. And that matters more than every feature combined.
If you are searching too
Start with the modality you most often fail to capture. For me, it was fleeting tasks. The ones that arrive while driving, while in a meeting about something else, while falling asleep. A shell alias solved it.
If your biggest loss is links, get Trove. The triage mode is genuinely satisfying. If you want a full GTD system and are comfortable with a database, Tandem GTD is the best implementation I have seen. If you want a universal capture front-end and do not mind processing later, DropMind is exactly what it claims to be.
But do not wait for the perfect tool. It does not exist. The tool that ships your thought from brain to trusted system in under five seconds is the right tool. Everything else is organization, and organization is what your AI agent is for.