Product·Code

The IDE for
agentic coding.

Code is Continuum’s primary surface - the workbench you live in. Spawn multi-provider agents into managed worktrees, steer with a real composer, review plan → diff → PR, and ship from Mac, web, Windows, Linux, or iPhone.

Download for Mac Read the docs
managed worktrees plan → ship multi-provider
UsageCodeSettings
Continuum
São Paulo · fable-5 · xhighworktree · sp-1 · Code
Migrate UsageHistoryLoader to actor isolation and keep the test suite green.
Migration plan4 steps
✓Parsers → nonisolated static
✓Cache writes on the actor
Bypass mtime for newest file
Re-run 800+ tests
continue with the cache path…
$0.41 · 48.2k tok · liveMac · web · iOS
01 · Managed worktrees

City-named sessions. Stable projects.

Each session lands in a managed worktree with a city name - São Paulo, Lisbon, Osaka - so parallel runs stay readable. Projects keep a stable order: oldest-first by first use, not “whatever just ticked.”

A managed worktree is a real git worktree: a second checkout of the same repository on its own branch, sharing one .git directory. Continuum creates it, tracks it, and cleans it up when you end the session. Four agents can edit four branches of the same project at once without one of them stashing another's work. If you want the raw mechanics, start with git worktrees explained and the Claude Code worktree flag.

Drag a project header to reorder; Continuum remembers. New repos append at the bottom. Within a repo, sessions still sort by recency so live work stays on top. Analytics follow the same canonicalization, so spend from a worktree rolls up to the repo it belongs to rather than fragmenting into a dozen paths.

city labels drag reorder managed only
Sessions across devices
Code
sidebar
search sessions…
⠿clawdmeter2
São Paulosp-1
Osakaos-2
⠿defx-frontend1
Lisbonlb-1
⠿continuum-web0
no sessions yet
oldest project firstdrag to reorder
02 · Composer

Built for how agents actually work.

Attachments, slash skills, @-mentions, voice dictation, and App Shots live in the same composer. Text and files survive send errors; a trailing newline keeps PTY agents happy.

Type / and the palette lists the skills actually installed for this session: global ones from your home skills directory plus any the repo ships, parsed from their own frontmatter. Codex sessions get their built-in command list instead. Typing @ opens a picker of open sessions, files the agent has already cited, and recent transcripts, so a reference is two keys rather than a copied path. The SKILL.md format guide covers what goes in one.

Swap model and effort mid-session without abandoning the thread. Toggle Plan vs Code permission modes. Cost and tokens tick in the footer while the turn runs, and a soft warning appears when you near a provider's weekly cap. Attachments stage into the session's own directory - inside the worktree when the agent is sandboxed to it - so a screenshot you pasted is a file the agent can actually open.

skills & mentions voice App Shots
Voice App Shots
Code
composer
Code Plan fable-5 xhigh worktree
screenshot.png spec.md
/review · project skill
@Sources/UsageHistoryLoader.swift run the migration plan and open a PR when green
~$0.12 · 14.6k tokCtrl+M voice · App Shots
03 · Plan to ship

Review is not a separate app.

Plan + live to-dos, Diff, PR create / approve / merge, Terminal, Sources, Artifacts, and Browser Preview with comment chips - all in the Code workbench. See the full flow on Plan & review.

The point is that reviewing is where agentic coding either pays off or quietly wrecks your repo. A tab-switch to a separate diff tool is a tab-switch you skip at 6pm. Here the diff is the same pane as the conversation that produced it, and every hunk is one click from staged, reverted, or committed. Our checklist for reviewing AI-generated code is the discipline these panes are shaped around.

01

Plan + to-dos

Step timeline from plan text. Approve & run when ready. Live checkmarks as work lands.

02

Diff

Live git diff, per-hunk stage / revert, commit from the same pane.

03

PR

Create, approve, merge. Status, checks, and +/− without leaving the session.

04

Terminal

Multi-pane PTY when you need raw output. Same session, same worktree.

05

Sources & artifacts

Cited files and agent-written PDFs / images with Quick Look preview.

06

Browser preview

Comment chips inject context back into the agent - selector + note, in thread.

Code
review
PlanDiffPRTerminalSourcesFiles
@@ UsageHistoryLoader.swift @@
- func parse(_ url: URL) {
+ nonisolated static func parse(
+ _ url: URL) async {
… 14 hunks · +212 −96
PR #214 ready to merge
+212 −96 · checks ✓
04 · Safety checkpoints

Every prompt has an undo.

Before a prompt runs, Continuum takes a checkpoint of the workspace. When the agent takes a wrong turn, you restore instead of archaeology through git reflog at midnight. This is a hard gate, not a setting you have to remember to switch on.

01

Plain folders too

A Local session can open any folder. If it has no .git, the create path initialises one rather than dead-ending your prompt. The base commit is empty, so your files stay untracked and ride a side snapshot instead of being staged behind your back.

02

Never your home

It refuses to initialise a repo at ~, above it, or in the standard containers like Desktop and Documents. A .git at home makes every git-aware tool treat your whole life as one repo. A project nested inside those folders is still fine.

03

Refusing never blocks

If a checkpoint genuinely can't be taken, you get one quiet notice and the prompt still goes. Only a transient failure worth retrying stops the send. Restore stays strict: it will not paper over a repo that moved.

Checkpoints run on the host, so a session driven from your phone against a Mac or an enrolled Linux box gets the same protection as one you started at the keyboard. Together with the Plan and Code permission poles below, that is what makes it reasonable to let an agent work unattended.

05 · Spawn mode

A grid of interactive CLIs.

Mac-only. Pick 4, 6, or 8 tiles and a mix - e.g. 4 Claude, 2 Codex, 2 Cursor. Continuum opens a terminal grid, all cwd’d to your home, for hands-on multi-agent work outside managed sessions.

Groups list under Spawns in the sidebar. They’re ephemeral - they die with the app, never persist, never pair to iOS. Expand any tile full-pane without tearing down the PTY.

Mac only 4 / 6 / 8 tiles ephemeral
Code
spawn 1
Claude 1› refactor parser…
Claude 2› waiting
Codex 1› apply_patch…
Cursor 1› idle
4 tiles · 2×2not paired to iOS
06 · Multi-provider

One workbench. Every plan you own.

Claude, Codex, Cursor, Grok, OpenCode, Antigravity, Z.ai, OpenRouter - each on its own login, no markup. Mid-session you can swap model and effort without abandoning the thread.

These are the vendors' real CLIs, not reimplementations. Continuum drives each one the way it wants to be driven and normalises what comes back into one transcript grammar, so a Codex tool call and a Claude tool call render as the same kind of row. That is why a Cursor session and a Claude Code session feel like one product instead of two embedded terminals.

Plan mode stays approval-gated and never auto-bypasses. Code mode skips approvals so the agent can ship. That two-pole model is deliberate: no half-trusted middle state where you think you're gated and aren't. See plan mode for the behaviour and Providers for per-vendor detail.

Claude Codex Cursor Grok OpenCode Z.ai OpenRouter
Code
session
Lisbon · defx-frontendmode · model · effort
Claude Codex Cursor Grok
Plan Code fable-5 high → xhigh

swap mid-session · Plan never bypasses

surfaces: Mac · web · desktop · iOSspawn grid: Mac only
07 · Questions

The honest fine print.

Short answers here, long answers in the docs.

Each Code session runs in an isolated git worktree Continuum creates and tracks - city-named, branch-scoped, and cleaned up when you end the session. You get parallel agents without stomping on your main checkout.

No. Spawn mode is Mac-only: interactive CLI tiles owned by the desktop app. Managed Code sessions - plan, diff, PR, composer - work on Mac, web/desktop, and iOS.

Plan stays approval-gated and never auto-bypasses permissions. Code skips approval gates so the agent can edit and run without a per-step confirm. You can still interrupt, swap model/effort, or switch back to Plan mid-session.

Mac is the reference host and UI. Web and the Windows/Linux desktop client render the same Continuum dashboard. iOS carries Plan / Diff / PR / Terminal / Files as a mobile workbench. Watch focuses on limits, plan approve/interrupt, and complications - not the full IDE.

As many as your machine and your provider limits allow. Each gets its own worktree and branch, so they don't collide in the filesystem. The real ceilings are CPU, your plan's rate limits, and how many diffs you can honestly review. See running Claude Code sessions in parallel for the practical numbers.

The composer footer ticks live cost and tokens for the turn in progress, priced per model. The Usage tab keeps the history by repo, provider, model, and day, parsed from the logs your CLIs already write. If you are on a subscription rail the marginal turn is prepaid, and the number reads as consumption rather than an invoice - tracking Claude Code cost explains the difference.

08 · Begin

Open Code.
Ship the branch.

Download for Mac, pair your phone, or open the web app, with the same workbench shape everywhere it matters. See every package on the Mac, Windows, and Linux desktop page.

Download for Mac Read the docs

related · providers · sessions · plan & review