Zed's new Delta app pairs an AI coding agent with Delta DB, a version control layer that links every code change to the chat that made it.
What is Zed Delta? #
Zed Delta is a standalone AI coding agent app from Zed, the team behind the Zed code editor. It’s separate from the editor itself and built around a different premise: instead of chatting with an agent in a side panel bolted onto a traditional IDE, you work almost entirely in threads, where a conversation with an AI agent is the primary interface for writing and changing code. Underneath that interface sits Delta DB, a new version control system Zed has been building that tracks every edit an agent makes and ties it back to the exact conversation that produced it.
The app is currently in early access, invite only, with sign-up open via email and GitHub username on Zed’s site. Pricing hasn’t been announced yet.
TL;DR #
Delta is a standalone agent-first app, not a plugin or panel inside the Zed editor, built around threads as the main way you interact with code.** Delta DB replaces commit-based history with continuous tracking**, recording every operation between commits and giving each one a stable identity linked to the conversation that caused it.Every line of code can be traced back to the chat that wrote it, and every conversation can be traced forward to the commits it generated.** Threads are multiplayer**, built on CRDTs, so teammates (and agents) can join the same thread via a link to co-edit prompts or continue work.** A browser version at delta.dev runs the same Rust app compiled to WebAssembly**, requiring no install and staying compatible with existing Git repositories.** Agents work in isolated clones of your repo**rather than your live working directory, using what Zed describes as a virtualized work tree that makes spinning up isolated copies cheap.Model access is flexible, supporting Zed’s own subscription, ChatGPT Plus/Pro sign-in, GitHub Copilot sign-in, and bring-your-own API keys for OpenAI, Anthropic, OpenRouter, and others.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does Delta DB differ from Git? #
Git works in snapshots. Every time you commit, Git freezes the state of your repository at that moment, but everything that happens between commits (every intermediate edit, every discarded attempt, every back-and-forth with an assistant) is invisible once it’s overwritten. Delta DB is designed to capture that in-between work instead of discarding it.
According to Zed’s pitch, Delta DB records every operation as it happens, gives each one a stable identity, and links it directly to the conversation that produced it. That means you can start from any line of code in a file and trace it back to the exact agent thread that wrote or modified it, and conversely you can open any thread and jump straight to the code changes it caused. Git treats commits as the unit of history. Delta DB treats the conversation as the unit of history, with commits acting more like checkpoints that fall out of that conversation naturally.
This is a meaningful shift in what counts as the “source” of a codebase. Zed’s argument is that if an AI agent wrote most of a feature, the conversation where you specified what you wanted, corrected its assumptions, and approved its output is arguably more informative than the diff itself. Delta DB treats that conversation as a first-class artifact worth preserving alongside the code.
How does the Delta app actually work? #
Opening Delta looks less like an editor and more like a chat app with a file browser attached. A sidebar lists projects and threads, and the main panel is a thread: a running conversation with an agent about a specific codebase. There’s still a real code editor available. Clicking a file in the tree opens it with full syntax highlighting, so you can inspect the actual code alongside the conversation that shaped it.
A demo run against a fresh Next.js scaffold showed the agent behaving in a fairly disciplined, structured way rather than just generating code on request:
- It read the README, package.json, and any project configuration files before answering questions, and noticed when documentation was stale compared to the actual code.
- When given an ambiguous instruction (paraphrased in the demo as “make it faster”), it stopped and presented multiple concrete interpretations to choose from, rather than guessing and running with one.
- Once given direction, it wrote a design specification and an implementation plan as actual documents in the repo, committing each one before writing code.
- It wrote tests before implementation, then patched configuration files to wire the new test script in.
- Each of these steps landed as a distinct commit, visible live in the thread as “external changes,” so the conversation and the Git history stayed in lockstep.
Notably, the agent did its work inside an isolated clone of the repository rather than the user’s live working directory. Zed’s framing is that Delta DB virtualizes the work tree, making these isolated clones cheap to create, so agents can experiment on branches without any risk of clobbering what’s on your screen. A branch switcher at the bottom of the app shows the main branch alongside the work-tree branches the agent spun up during the session.
Is Delta worth trying right now? #
It depends on what you’re looking for. Delta is early access software, invite-gated, with no public pricing yet, so it’s not a tool you can casually adopt for a production team today. The core idea, tying every code change to the conversation that produced it, is distinctive among AI coding tools, most of which are still built as chat interfaces layered on top of a conventional diff viewer. Delta instead rebuilds the version control layer itself around conversations, which is a deeper and more structural bet than most competitors are making.
The multiplayer angle is also worth noting for teams. Because threads are backed by CRDTs, multiple people, and multiple agents, can be present in the same thread simultaneously, commenting, editing prompts, or picking up work someone else started. Sharing a thread produces a link a teammate can open to see the full history and continue talking to the same agent, which is a different collaboration model than passing around a pull request.
Model flexibility is a practical plus. Rather than locking users into one provider, Delta supports Zed’s own model access through subscription, direct sign-in with a ChatGPT Plus or Pro account to use OpenAI models, a GitHub Copilot sign-in option, and the ability to plug in personal API keys for OpenAI, Anthropic, OpenRouter, and other providers. The demo showed 24 models available in the picker.
The tradeoffs are real, though. Working in threads instead of a traditional editor is a workflow change, and not everyone wants to give up a familiar IDE-centric setup for a conversation-centric one. It’s also unproven at scale: an early access demo on a fresh scaffold project doesn’t tell you how Delta DB performs on a large, messy, years-old codebase with dozens of contributors.
What’s the pricing and availability? #
Delta is in early access with invites rolling out gradually. There’s no published pricing yet. Interested users can sign up with an email address and GitHub username through Zed’s site. Zed has stated it intends to open source Delta DB, with optional paid services layered on top, and eventually integrate Delta DB’s capabilities into the main Zed editor as well.
Frequently Asked Questions #
What is Zed Delta?
Zed Delta is a standalone AI coding agent application from Zed, separate from the Zed code editor, built around thread-based conversations with AI agents as the primary way of writing and modifying code.
What is Delta DB?
Delta DB is Zed’s new version control layer. Unlike Git, which only captures snapshots at commit time, Delta DB records every operation between commits, assigns each a stable identity, and links it to the conversation that produced it, so code changes and chat history stay connected.
Is Zed Delta a replacement for Git?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
It’s built to be a layer that sits alongside and extends Git-style version control rather than a wholesale replacement. The browser version at delta.dev stays compatible with existing Git repositories, so collaborators who aren’t using Delta won’t notice any difference in the underlying repo.
Can I use my own AI models with Delta?
Yes. Delta supports Zed’s own subscription-based model access, direct ChatGPT Plus or Pro sign-in, GitHub Copilot sign-in, and personal API keys for providers including OpenAI, Anthropic, and OpenRouter.
Is Zed Delta available to the public?
Not yet in full. It’s currently in early access with invites rolling out, and pricing hasn’t been announced. Sign-up is available through Zed’s website using an email address and GitHub username.