Comparisons

Codecast vs Delta

Delta, from the makers of the Zed editor, is a new app where a thread owns its own checkout and records every edit between commits. Codecast records and steers the agent sessions your team already runs, in the terminals and on the machines you already use.

Codecast is a team record and control layer for coding agents: a daemon on each machine watches the history files Claude Code, Codex, Cursor and Gemini already write and syncs every session into one searchable, steerable place, with tasks, plans, docs, triggers and line attribution built on that record.

Delta is a desktop app from Zed Industries, in public beta since September 2026, built around the thread: a conversation with Delta's own agent, paired with an isolated checkout of your repository. Its sync layer, DeltaDB, records each edit and message in order between git commits and replicates the conversation and the working tree to everyone in the thread in real time. Delta’s site.

Bottom line

The two products make opposite bets. Delta asks you to move the work into one new app so it can own the checkout, run the agent and replay every keystroke of the change to your teammates. Codecast asks you to change nothing about how you run agents and gives your team the record, the memory and the controls around what those agents did. Pick Delta when the live, edit by edit, multiplayer thread is the thing you want. Pick codecast when your team already runs several agents on several machines and needs to see, search, steer and attribute all of it.

Side by side

CodecastDelta
Core modelRecords and steers the agent sessions you already run in your own terminals; adds a team layer on topA new app with its own agent; each thread owns an isolated checkout and DeltaDB records every edit between commits
Agents supportedClaude Code, Codex CLI, Cursor, Gemini (OpenCode and pi in progress)Delta's built in agent on Anthropic, OpenAI, GitHub Copilot, Grok, OpenRouter and other providers; a plugin that syncs Claude Code terminal sessions into threads is on the roadmap as in progress
Where the agent runsIn your own terminal, on whichever machine you started it; codecast never runs the model itselfOn the sender's machine from the desktop app; a turn sent from the browser uses a Delta hosted model and has no shell or local tools
Unit of workThe session; tasks, plans, docs and projects link to the sessions that did the workThe thread: one conversation, one Delta worktree per project, one checkout per participant
What gets recordedThe agent's own transcript: every message, tool call, permission prompt and file edit, plus the commits that followedEvery edit and message in sequence in DeltaDB; the repository's git objects and file contents are uploaded so the thread can be replayed anywhere
IsolationYour checkout by default; cast ws creates a git worktree with its own env files, ports and setup when you want isolationIsolated Workspace by default: a managed checkout per thread under .delta; you can point a thread at your existing folder, and then threads that share it are not isolated
ReviewQuote and comment on any reply, leave notes on diff lines, and send them as one batched review; review a teammate's session from its transcript and diff; batch reviews on pull requestsInline comments anchored to code passages in the Changes tab; review subthreads with an agent written change guide and an Approve or Request Changes verdict; Land Changes runs a landing skill in a subthread
Team visibilityEvery teammate's sessions in one feed and inbox, with per-directory privacy controlsThreads are private until shared: invited by email (ten per thread), organization wide, or any signed in user with the link
Working in one thread togetherAnyone with access reads the live transcript and can message the session; replies carry the sender's name; a session can have several owners and sit in several inboxesEveryone sees each other's drafts as they type; all pending drafts go to the agent as one turn; the sender's account pays for the turn and it runs on the sender's machine
Search & memoryFull-text search across all past sessions; agents query team history themselves (cast search / cast ask)Find within a thread; archived threads stay searchable; no documented search across threads and no memory the agent consults
Line-level attributioncast blame traces any line of code to the conversation that wrote itNot documented as a feature; Delta's references anchor to changes rather than line numbers
Remote steeringWeb, desktop, and iOS apps; answer permission prompts from any devicedelta.dev/threads in any browser, phones included, for reading, commenting and reviews; creating threads and running terminals need the desktop app
AutomationTriggers on a delay, an interval or a webhook; workflows with human gates; plans that run tasks in waves; org roles with standing agent sessionsSubagents with Scout, Worker and Reviewer profiles in TOML (four per thread and eight overall by default); a landing workflow per project
Agent safetyThe agent's own permission prompts stay in force and you can answer them from web, desktop or phone; secrets are redacted before syncNo permission prompts and no sandbox today: the agent calls tools without asking, and prepare scripts, direnv and AGENTS.md run before you review them; secrets are redacted locally before upload
Where your data livesA backend you can host yourself; project paths are hashed, each conversation has its own privacy level, and an optional mode encrypts bodies so the server cannot read themZed's servers on Cloudflare (R2, Durable Objects, KV, D1), encrypted at rest with Cloudflare managed keys; no self hosting; deleting a thread locally leaves the server copy until you delete the account
PlatformsDaemon on macOS, Linux and Windows; web, a macOS desktop app, an iOS app, and blame in VS Code, Cursor and vimmacOS 13 or later on Apple Silicon, Linux and Windows; one app instance at a time; a browser viewer
Sign inEmail and password or GitHubGitHub through a Zed account; no organization sign on documented
StatusGenerally available; free for individuals, per seat for teamsPublic beta; free during the beta, with paid plans announced as coming on the Zed editor's billing
Open sourceMIT, self-hostable backendClosed source, built by Zed Industries

Strengths and weaknesses

Codecast

Strengths

  • Records every session with no change to how anyone works: four agents, any machine, nothing to switch on.
  • The record outlives the session. Full text and semantic search, questions across the whole corpus, and cast blame from any line back to the conversation that wrote it.
  • Agents read the record themselves, so memory carries across sessions, people and months.
  • Steer from anywhere: answer permission prompts, send instructions, fork, restart and reassign sessions from web, desktop or phone.
  • Work tracking grows out of the record: tasks, plans, docs, projects, a decisions queue, triggers, workflows and org roles all link back to sessions.
  • MIT licensed with a backend you can host yourself; per conversation privacy and optional encryption the server cannot read.

Weaknesses

  • It depends on each agent's own transcript format, so what a client does not expose stays absent: Cursor sessions cannot be launched or messaged, and Cursor and Gemini have no live state classifier.
  • The checkout is shared unless you ask for a worktree, so two sessions in one folder can collide.
  • Teammates read the transcript and the diffs; they do not get a live synced copy of the working tree to edit in the browser.
  • Comments attach to messages and file lines, not to a passage that follows the code as it moves.
  • It is not an editor. You still write code yourself in your terminal or IDE.
  • A large surface. The inbox, tasks, plans, docs, chat, roles and triggers take time to learn.
Delta

Strengths

  • DeltaDB records every edit between commits and replays the conversation and the working tree together to every participant in real time.
  • Each thread starts in its own checkout, so parallel threads cannot trample each other or your working folder.
  • Review happens where the work happened: comments anchored to passages, review subthreads with a change guide, an approve or request changes verdict, and a landing skill that ships the change.
  • Multiplayer composing: drafts are visible as people type and go to the agent as one turn.
  • One agent, many providers. Switch models mid thread, sign in with ChatGPT, Copilot or Grok subscriptions, or bring API keys.
  • A browser viewer for reading, commenting and reviewing with nothing installed, phones included.

Weaknesses

  • Public beta. The Claude Code plugin, MCP support, a remote runtime, sandboxing, permission prompts and conversation branching are roadmap items, not shipped.
  • The agent acts without asking. Prepare scripts, direnv and AGENTS.md run before you review them.
  • Your repository's file contents and git objects are uploaded to Zed's servers; there is no self hosting, and deleting a thread locally leaves the server copy.
  • Sessions you run in your own terminal are not recorded today. The work happens inside Delta's app and Delta's agent.
  • Real work needs the desktop app: macOS requires Apple Silicon, only one app instance runs at a time, and browser turns have no shell.
  • Subthreads go one level deep, keybindings cannot be rebound, and there is no documented search across threads or memory for the agent.

Two bets on where the work lives

Delta's founding claim is that the thread is where software happens now, so the thread should be the workspace. You add a project, Delta imports its files into DeltaDB, and every thread you open gets its own checkout under .delta with the conversation and the edits recorded together. Zed says it has turned pull requests off on Delta's own repository and lands changes from threads instead.

Codecast starts from the opposite observation: your team already runs agents in terminals, IDEs and tmux panes on many machines, and the problem is that all of that evaporates when the pane closes. So the daemon records what is already happening and the product grows around the record: an inbox of every session by who acts next, search and memory over the corpus, and a control channel back into any live session.

The practical difference is what you give up. Delta asks for a new app, a new agent and an upload of your repository. Codecast asks for a daemon and leaves the agent, the terminal and the git workflow exactly as they were.

What each one records

Delta records operations. Each edit gets a stable identity and a place in sequence, so you can scrub a thread back to any point, revert the conversation to the cursor, comment on a passage while the agent is mid turn, and hand a teammate the exact state of the working tree without a commit. That is a stronger primitive than a diff, and it is the reason Delta had to be a new application rather than a feature of git or of Zed.

Codecast records the agent's own transcript: every message, tool call, permission prompt, thinking block and file edit, with diffs materialized per edit and commits tied back to the session. It does not replay keystrokes, but it keeps things Delta does not: which commands the agent ran, what it read, which prompt it was answering, and what the person said back. cast blame then joins git blame to that record so any line resolves to the conversation and the message that wrote it.

The upload differs too. Delta stores your repository's git objects and file contents on Zed's Cloudflare infrastructure so that browser participants can see the worktree. Codecast stores transcripts and diffs, hashes project paths, redacts secrets before sync, and can run on your own backend.

Isolation and git

Delta isolates by default. Every thread gets a managed checkout, every participant gets their own checkout of that worktree on their own machine, and a prepare script runs when a checkout is created. The agent can push to a second remote named local, which points at your own repository, so a teammate can pick up a branch without going through origin. The cost is that two threads on the same branch name are two separate checkouts, and uncommitted work in one does not appear in the other.

Codecast works in your checkout by default and treats isolation as something you ask for. cast ws acquire creates a git worktree with copied env files, its own port allocations and the project's setup commands, and cast ws destroy tears it down; a warm pool can keep worktrees ready. Because sessions run real git on real branches, landing work is ordinary: commit, push, open a pull request, and codecast binds the session to that pull request so reviews and failing checks wake it.

Review

Delta's review is the strongest part of the product. The Changes tab shows diffs against the branch base, the last commit or the last turn. You select a passage and type to comment; comments queue until your next message and reach the agent as targeted feedback on that exact passage. A review subthread asks the agent for a change guide, a walkthrough of the change in the order that explains it, and ends with Approve or Request Changes posted back to the parent. A skill marked as a landing action turns the merge into a subthread that runs checks and publishes.

Codecast reviews at three levels. In a conversation you quote any reply, leave notes on diff lines and send them as one batched review to the session. Across sessions, a reviewer reads a teammate's transcript and diff and reports back to that session or its owner. On GitHub, cast pr holds notes and submits them as one review with a verdict, and the session that owns the pull request is woken with the whole review as one message.

Working together

In a shared Delta thread everyone can message, steer, edit files and comment, and drafts are visible as people type. All pending drafts go to the agent as one turn. The turn runs on the sender's machine with the sender's tools and credentials and is billed to the sender, so what the agent can reach depends on who pressed send. Browser participants get replicated files and a hosted model, not a shell.

Codecast's collaboration is built around the inbox rather than the thread. Each session has one or more owners and appears in their inboxes sorted by who acts next; passing a session moves it to a teammate; cast send messages any session and the reply carries your name. Around the sessions sit team chat, transcribed calls with action items, a decisions queue that lets an agent ask a question without stopping, and an org model where roles hold standing sessions and a line of stations carries a task from analysis to review.

Safety and where the data goes

Delta's own docs are direct about this: the agent acts autonomously and does not ask before calling tools, there is no sandbox, and Delta runs a repository's prepare script, direnv file and AGENTS.md without asking you to review them first. Permissions, sandboxing and worktree trust are on the roadmap. Data is encrypted in transit and at rest on Cloudflare with Cloudflare managed keys, secrets are redacted on the device before upload, and account deletion is by email.

Codecast leaves the agent's own permission model in place and lets you answer prompts from any device. It redacts secrets before sync, hashes project paths, supports a private, summary only or full visibility level per conversation, and can encrypt conversation bodies so the server never reads them. The backend is MIT licensed and documented for self hosting.

What is not there yet

Delta's roadmap, as of this reading, lists the Claude Code plugin, MCP support, a remote runtime, repository based access and web parity as in progress, and mentions, repository level context, sandboxing, WSL, conversation branching and a graph view as up next. Until the plugin ships, Delta only records work done through Delta's agent.

Codecast has its own gaps, stated in its own docs. OpenCode and pi support is in the tree but pre release. A Codex fork inherits its parent's history but cannot yet take a follow up turn. Cursor sessions are ingested and resumable but cannot be launched or messaged from codecast. The team tier is early access with no self serve billing yet.

Choose Delta when

  • You want the thread to be the workspace: an isolated checkout per conversation, every edit replayed live to teammates, and review anchored to passages of the change.
  • You are happy to run one new app with its own agent, and to upload the repository so that anyone with a browser can follow along.
  • You want to land changes from a thread instead of a pull request, and your team is small enough to invite by email.

Choose Codecast when

  • Your team already runs Claude Code, Codex, Cursor or Gemini on several machines and you want all of it recorded, searchable and steerable without changing how anyone works.
  • You need the record afterward: search months later, agents that consult team history, and a line of code traced to the conversation that wrote it.
  • You want the agent's permission prompts to stay in force and to answer them from your phone.
  • You need the data to stay on infrastructure you control, or you need tasks, plans, triggers and roles built on the same record as the sessions.

Today they do not overlap on disk: Delta's agent runs inside Delta and writes no transcript for a codecast daemon to watch. If you run Claude Code in a terminal against the same repository, codecast records that session as usual. Once Delta's Claude Code plugin ships, the same terminal session could sync into a Delta thread and into codecast at the same time.

Delta facts read from its public docs on 2026-09-20. Tell us if something changed.

Try codecast on the sessions you ran today