claude, codex, agy, cursor-agent, grok, and opencode. Each CLI runs as a child process on your machine. Continuum reads the local state those CLIs already maintain, spawns and steers their sessions, and surfaces their usage — it does not proxy or inspect their model network traffic.
A ninth lane, Continuum-hosted inference, needs no local CLI: Continuum serves the models itself and bills them to your Continuum account instead of your own provider subscription. It appears as a Continuum rail in the model picker, split into Included in Free — a roster of free models available on every plan, including no plan at all — and the frontier models a paid tier unlocks. Because the roster is served rather than bundled, read the live tray for what is currently offered; the hosted inference page covers billing and org policy.
Provider summary
- Transport is the runtime path the daemon uses. Claude defaults to the Agent SDK’s ACP adapter and switches to a real per-session PTY when you pick the CLI backend under Settings → Advanced → Claude runtime — that PTY is the only direct-PTY path in the app. Codex runs through
codex app-server; Cursor runs through the native ACP harness; Grok uses ACP (grok agent stdio) when the installed CLI advertises it and falls back to its headless one-shot transport when it doesn’t; Antigravity/Gemini runs through its own transport-owning headless harness. OpenCode Go, OpenRouter, and Z.ai Coding are three branded lanes sharing oneopencode serveprocess. - Quota gauge means Continuum polls a live usage source for that account and shows a rolling-quota meter in the Usage tab — and, for every provider except Z.ai Coding, in the Mac menu bar too. OpenCode Go exposes 5h / weekly / monthly windows, but only through its web dashboard — you supply a workspace id and session cookie under Settings → Providers → Configure quota tracking; there is no API-key usage endpoint yet. OpenRouter is bring-your-own-key with no subscription quota at all, so it is tracked as a dollar-cost lane computed from token counts and the bundled pricing snapshot. See the OpenRouter page.
- Effort dial is a reasoning-effort control that actually reaches the CLI. Three providers accept one, and each accepts a different vocabulary: Claude takes
--effort low|medium|high|xhigh|max(plus the Claude-only Ultracode mode, set through a session settings key rather than a flag), Codex takesmodel_reasoning_effortfromminimalthroughmax, and Grok takes--reasoning-effort low|medium|high. The composer’s dial only offers the rungs the selected provider accepts, and a value carried across a provider swap clamps to the nearest supported rung at the spawn boundary — so the CLI never sees a level it would reject. Providers not listed here select a model but pass no effort value.
Auto model routing
At the top of the model picker, above every provider rail, sits Auto — “Picks the best models for the task.” Choosing it hands model selection to Continuum’s router instead of pinning the session to one model. Auto is a choice you make when a session starts, not a background behavior. It is not on by default, and it is not a mid-session Change model action: an already-running session keeps the model you gave it. What Auto replaces is the one-model assumption, not your control. An Auto run is a state machine on the host, and you watch it move through its phases:1
Triage and route
The router reads your request first. A small one takes a fast path — one executor, no ceremony. Anything substantial gets the full pipeline: a planner chosen for raw capability, an executor chosen for capability plus agentic strength (and drawn from a taste-weighted pool when the task is design- or copy-shaped), and a verifier deliberately taken from a different provider where one is available, so no model checks its own output. Each assignment carries a written reason you can read — down to “no taste ≥ 7 model available” or “same-provider fallback”.
2
Plan, then wait for you
The planner produces a plan and the run parks in awaiting approval. Nothing is written until you press Approve & run. Cancel is the brake, at this or any later phase.
3
Execute, verify, fix
The executor does the work, the verifier checks it, and a fix pass runs if verification finds problems. Every stage is stamped with the model, provider, and account that actually ran it — including a re-stamp when a model is unavailable and the run falls back down its chain — so a finished run reads as receipts, not a guess. Live cost accrues per stage.
Auto needs a host that serves the router. Macs and the Windows/Linux/CLI Go agent both do on current builds; against an older host the Auto entry doesn’t render at all rather than failing at send time. Runs survive a host restart — an interrupted run suspends and can be resumed rather than starting over.
Auth stays with each tool
Continuum has no model credential of its own. Where a provider signs in through its CLI (Claude, Codex, Antigravity, Grok), Continuum reads the credential that CLI already stored and never runs a login that could clobber it. Where a provider authenticates with an API key (Cursor, OpenCode Go, OpenRouter, Z.ai Coding), pasting the key in Settings → Providers writes it into that tool’s own credential file — for the OpenCode-connector lanes that is~/.local/share/opencode/auth.json, the same file opencode auth login writes.
The one credential Continuum keeps a copy of is Claude Code’s OAuth access token, imported into Continuum’s own Keychain entry so the rate-limit gauge can poll without the CLI running. Where Continuum reaches a provider’s billing/usage endpoint for a quota gauge, it authenticates with that stored credential, makes no destructive calls, and keeps the result on-device. The Security and Privacy pages describe exactly what leaves your machine.
Multiple accounts are supported on every provider. You can sign in more than one subscription or key and run them side by side, pinning sessions to a specific account. Each non-primary account gets a real isolation boundary: its own config root applied at spawn time (
CLAUDE_CONFIG_DIR for Claude, CODEX_HOME for Codex, XDG_DATA_HOME for the OpenCode-connector lanes, GROK_HOME for Grok, an isolated HOME for Antigravity because agy reads .gemini from HOME and nothing else; Cursor has no documented config-root override, so its extra accounts are keyed on an explicit CURSOR_API_KEY instead), a per-account Keychain partition for Claude tokens, and an environment scrub so one account’s credentials can’t bleed into another’s spawn. HOME is otherwise left untouched, so git, ssh, gh, and node keep working inside worktrees. Add an account in Settings → Providers → “Add account…”; the account picker appears once a provider has two or more accounts. See multi-account.The account index lives at
~/Library/Application Support/Clawdmeter/provider-instances.json — no secrets in that file. Clawdmeter is the app’s internal codename and still names its on-disk paths.Provider pages
Claude Code
Agent SDK and direct-PTY backends, Plan/Code permissions, the effort dial and Ultracode, rate-limit gauges, and usage analytics from JSONL.
Codex
App-server harness, sandbox-gated approvals, the full minimal-to-max effort ramp, and respawn-lineage transcript resolution.
Antigravity / Gemini
Headless
agy harness with conversation-DB analytics, brain-dir plan snapshots, and a three-tier quota probe.Cursor
ACP harness sessions via
cursor-agent, account-scoped model discovery, and a Cursor billing-period panel in Usage.Grok
ACP or headless
grok sessions, a three-rung effort dial, and a live weekly credit gauge.OpenRouter via OpenCode
opencode serve SSE integration with a bring-your-own API key and dollar-cost usage tracking.Z.ai Coding
GLM Coding Plan models through the OpenCode connector, with a 5h + weekly quota gauge and its own analytics column.
Continuum hosted
Pooled inference billed to your Continuum subscription — no provider account and no local CLI required.