Zed Delta: The Public Beta That Replaces Pull Requests with Agent Threads

Zed Delta is a standalone multiplayer environment for coding with agents and reviewing what they build, from the creators of the Zed editor - available at delta.dev with native downloads for Windows, macOS, and Linux, a no-install web version, and mobile-browser thread access. It entered private beta on August 12 2026 and opened as a free public beta on September 16 2026. The unit of work is the thread: you invite teammates directly into your agent conversations, they see the same worktrees on their own machines, and they can pick up the thread where you left off. Comments attach to the conversation, a diff, a plan, the thinking, or any worktree line - anchored to the operation, not a stale line number. It is deliberately not a Zed editor feature drop: the editor keeps shipping separately while DeltaDB gets polished in this dedicated client.

During the public beta, yes - Zed's announcement states Delta is free during the public beta with no time limit published. Paid plans for individuals and teams are announced but carry no prices or GA date, and Zed commits that there will always be a free version of Delta. Hosted models are planned to be token-metered with no API keys required on paid plans, but usage economics are unpublished in beta. The zero-extra-cost paths today: connect a subscription you already pay for (ChatGPT Plus/Pro, GitHub Copilot, or an eligible Grok/X plan), or bring your own API keys for Anthropic, OpenAI, OpenRouter, OpenCode Zen, and OpenCode Go - keys read from Delta's config .env on Windows at %USERPROFILE%\.config\delta\.env.

The argument: since GitHub introduced pull requests over 15 years ago, review has meant commenting on a diff after the commit - and with agents generating more code than ever, the diffs mushroom while the decisions behind the code stay invisible. A reviewer feeding your diff into another agent has to piece together context you already worked through. Delta starts review from the conversation instead: invite anyone into the thread where the work happened, or open a dedicated review subthread with its own isolated copy of the parent thread's worktrees, so reviewers can explore code and try changes with agents without disrupting the original work. If a reviewer spots a problem they can request a revision or fix it with the agent themselves - fixes fold back into the parent thread before you ask the agent to land the change. Zed is proving the workflow on itself: pull requests are disabled on Delta's own repository, where 33 people have landed 570 changes to main entirely in threads.

DeltaDB is the new version-control layer underneath Delta, built on one idea: software is made between commits. Where Git captures a snapshot at each commit, DeltaDB captures every operation in between - each edit, each human and agent message - gives each one a stable identity, and records the message and the edit it produced side by side so neither drifts from the other. Every reference anchors to a delta instead of a line number, so it survives as code moves: from any line of code you can find the conversation that produced it, and from any point in a past conversation you can jump to that code now or as it stood when written. Worktrees are conflict-free replicated (CRDTs), so many people and agents can edit the same files at once across machines. Commits remain the checkpoint you push, pull, and build from - DeltaDB retains the work between them - and existing Git repositories work unchanged: teammates who never open Delta see a normal Git repo.

Three ways in, per Delta's own documentation. Hosted models on paid Delta plans, metered on a token basis with no API keys required. Subscription sign-ins: ChatGPT (Plus or Pro) for OpenAI models, GitHub Copilot via your GitHub account, and Grok via an eligible Grok or X subscription - ChatGPT and Grok connections sync through your Delta account; Copilot credentials stay on the connecting device. Bring-your-own API keys for Anthropic, OpenAI, OpenRouter, OpenCode Zen, and OpenCode Go, added in Delta's settings or read from ~/.config/delta/.env (on Windows, %USERPROFILE%\.config\delta\.env); keys exported in your environment take precedence and show as read-only. OpenRouter additionally shows a curated set of models Delta has evaluated for agent work, and you can add exact OpenRouter model IDs - those arrive marked not evaluated.

Yes - the public beta ships native downloads for Windows, macOS, and Linux, plus a web version at delta.dev that runs without installing anything, and threads stay reachable from a mobile browser. Windows API keys land in %USERPROFILE%\.config\delta\.env. The web app is the same Rust client compiled to WebAssembly, so the experience matches the desktop rather than being a stripped-down viewer. Git compatibility is unchanged: point Delta at an existing repository and teammates who never install Delta see a normal Git repo - adoption can be per-thread, not per-team.

They solve different halves of the review problem. Delta keeps the conversation and code together: review starts before a commit exists, with the agent's decisions attached, and reviewers can join mid-run. Pull requests keep the world's ecosystem - issues, CI, community contributions, the place every open-source project lives. Zed's own sequencing shows the coexistence plan: zed-industries/zed stays on GitHub for issues and contributions while contributors share Delta threads alongside their pull requests, and GitHub remains the system of record for people who never open Delta. The honest decision is not keep-Git-or-leave-it: it is whether your team's unfinished, pre-commit conversation belongs in a shared store. If review is your bottleneck because agents generate more than humans can diff-review, the thread model is a real upgrade; if you are a solo developer or PR-light team, a multiplayer review client buys you little today.

Yes - and they solve different layers. VibeFuse is the first free widget-based AI harness: it runs the official Claude Code, Codex CLI, Gemini CLI, Cursor, and Grok Build CLIs - plus any Model-API agent - as widgets over your real repositories on Windows. The harness costs \$0 with no usage meter: your vendor accounts do the billing, and you switch models or vendors per task instead of being locked to one meter. When a quota wall or rate limit hits one widget, route the task to another vendor's CLI on the same canvas. Delta is a multiplayer review-and-collaboration layer; VibeFuse is the execution canvas where the agents themselves run. They compose: run agent work through VibeFuse widgets, review it in Delta threads. VibeFuse also runs an open-source marketplace where creators sell widgets, skills, and styling packs for revenue share - pair it with VocalFuse for local voice dictation while you build.