- Hosts run the agents. A host has your code on disk, your provider CLIs
installed, and a local daemon that supervises every session. That’s the macOS
app, the Windows and Linux desktop apps, the
continuumCLI, a VPS, a Tailscale box, or a machine Continuum provisions for you. - Remotes steer them. iPhone, Apple Watch, Android, and the web app hold no code and no provider credentials — they connect to a host and drive it.
The three jobs
- Run and control coding-agent sessions. Claude Code, Codex, Antigravity/Gemini, Cursor, Grok, OpenCode-connected models (OpenRouter, Z.ai Coding), and Continuum’s own hosted models — in isolated git worktrees, with plan review, diffs, PRs, and terminals. Pick Auto instead of a model and Continuum routes the work across your connected providers itself. This is the Code tab, and it’s the center of the product.
- Show live quota and spend. Session windows, weekly caps, reset timers,
and historical
$ / tokenanalytics by provider, day, and repo. This is the Usage tab. - Follow you off the desk. Sign in on your phone, watch, or a browser and every session is there: approve a plan, send a follow-up, star the better answer, merge a PR. End-to-end encrypted between your devices.
Continuum shipped a Chat tab and a Cowork tab earlier in its life. Both
are hidden by default as of 2026-07-30 while the product focuses on Code. The
docs for those surfaces are still here — see Chat — but you
won’t find the tabs in a current build.
What it costs
Every client is free: Mac, Windows, Linux, CLI, iPhone, Watch, Android, web. You bring the provider subscriptions you already pay for, and Continuum drives each provider through its own official CLI, so the work lands on your flat-rate subscription rather than a metered API key. You can sign in more than one account per provider and run them side by side — see multi-account. If you’d rather not bring keys at all, Continuum-hosted inference is the one paid add-on: a monthly plan with a weekly inference budget, starting at $25/mo. A set of free hosted models is included on every plan — the model picker’s Continuum rail splits them into Included in Free and frontier models — so you can get a real answer before connecting anything. See hosted inference.The shape of the thing
A host runs an HTTP/WebSocket daemon (default port21731, WebSocket on the
next port up) and the provider CLIs as child processes. Remotes connect to it
over an end-to-end encrypted relay — the relay forwards sealed bytes and cannot
read them.
Continuum is local-first. Code, transcripts, analytics, and provider
credentials live on your hosts; remotes hold a token and a view. The
privacy doc enumerates exactly what can leave your machines and
why.
Pick a path
Quickstart
Install on any platform, light up your first gauge, and run your first
session in five minutes.
Installation
Per-platform requirements, installers, and how each client updates itself.
Connect your devices
Sign in with the same email on each device and approve it once. No QR code.
Windows & Linux desktop
The same workbench outside Apple’s world, with its own bundled daemon.
Web app
The full control plane in a browser tab — nothing to install.
Terminal CLI
continuum drives the same sessions from a shell, a script, or CI.The Code tab
Sessions, worktrees, plans, diffs, and PRs — where the work happens.
Providers
What Continuum needs from each CLI, and what lights up when it finds it.