Know what's left on your coding plans before you dispatch

See how much is left on your coding plans before your agents start new work.

Quick start | Works with | Dashboard | CLI reference

The problem

You run an orchestrator that fans work out across models on different accounts: Kimi for one task, GLM for the next, MiniMax for a third. One of those accounts runs out of quota. The orchestrator can't tell, so it keeps sending sub-agents there, and each one fails on a quota error instead of doing the work.

Claude Code shows its own rate-limit state, but nothing shows the others in one place. Every provider keeps its numbers in its own console, in its own shape: a rolling 5-hour window plus a weekly one for MiniMax and Kimi, a percent used for GLM, a balance for DeepSeek and OpenRouter.

The solution

view-limits asks each provider how much is left, normalizes the answer to one shape (a state, the windows or balance, and the reset time), and caches it in a local status.json. /view-limits renders that as a live view. One CLI, vl.js, does all of it, so Claude Code, Codex and OpenCode show the same thing.

There's an optional gate too. A PreToolUse hook resolves each sub-agent dispatch (Agent/Task, and Traycer's create_agent, configure_agent and fork_agent) to a route and can deny it when that account is exhausted. It's off by default, and it's built not to be wrong in the annoying direction:

• It reads the cached status.json synchronously and never calls a provider API.
• It denies only a route whose status is fresh and unambiguously exhausted. Unknown, stale, unmapped and ambiguous routes are allowed.
• Anything unreadable in its own config counts as off.

An optional dashboard daemon keeps the status fresh in the background and serves a page on 127.0.0.1, with an opt-in, read-only view for your phone. Built-in provider rows show the provider's logo (bundled with the plugin, nothing is fetched), and custom providers get a generic icon. /view-limits:dashboard brings it up: it starts the daemon if it isn't running, reuses it if it is, and opens the page in your browser, so you never have to go to a terminal to look at your limits.

Your keys stay out of chat. /view-limits:setup opens a local browser form, and the keys go to the macOS Keychain, or to an AES-256-GCM encrypted file on other systems.

Quick start

1. Add the marketplace and install the plugin:

   claude plugin marketplace add KHAEntertainment/marketplace
   claude plugin install view-limits@kha-marketplace

2. Run /view-limits:setup. It opens a local browser form for the providers you want to monitor. Paste your keys there, not in chat. A fresh install has no providers until you add them (it imports a MiniMax key from ~/.mmx/config.json if it finds one).

3. Run /view-limits.

You'll see one block per route you added: its state, the quota windows or balance, and when it resets. A route whose numbers can't be proven fresh shows unknown, never healthy.

Optional: run /view-limits:dashboard to keep that view in a page instead. It starts the background daemon if it isn't running, reuses it if it is, and opens http://127.0.0.1:<port>/ in your browser. /view-limits:dashboard-stop shuts it down and /view-limits:dashboard-status reports what it last collected; vl.js dashboard open --no-open prints the address and opens nothing. The three commands take no arguments — a command that passes one would be handing shell text to a pre-approved call.

Nothing is blocked yet. The gate starts off, and the cache refreshes when you run /view-limits (set refreshOnSessionStart to warm it at session start). To get advice in the model's context, or to turn blocking on, see Hooks.

Codex and OpenCode run the same vl.js view command: Codex setup, OpenCode setup. To run from a checkout, use claude --plugin-dir /path/to/view-limits. More in Install.

Every command here runs its CLI without asking you first: each one declares allowed-tools: Bash("${CLAUDE_PLUGIN_ROOT}/bin/vl.js" *), which pre-approves this plugin's own vl.js and nothing else. Claude Code checks a command file's injected !`…` call against your permission rules while the file renders, and outside auto mode anything that is not allowed aborts the whole invocation, so the grant is what makes /view-limits print instead of failing. One honest caveat: the docs state that substitution inside allowed-tools for a plugin skill, and a plugin command file accepting the same field is documented adjacent to that, not spelled out — the first real session settles it.

Those commands take no arguments, and that is deliberate rather than an oversight. A ! exec is shell text, and Claude Code substitutes $ARGUMENTS as typed, so an argument in one of those files would be code running under the very grant above — not data for the CLI to read. So the surfaces that need a second verb got a second command (/view-limits:dashboard-stop, /view-limits:dashboard-status), and the things that genuinely take a route id are terminal commands: vl.js setup <route-id>, vl.js update <route-id>, vl.js dashboard lan enable --interface <ip>. test/relay-only.test.js and test/shim-allowed-tools.test.js pin the one fixed call per file and reject any substitution but ${CLAUDE_PLUGIN_ROOT}.

Each of those injected calls also carries --relay, because Claude Code aborts a command whose injected exec exits non-zero — and "the dashboard is not running" is a normal answer, not a failure (dashboard status exits 3, dashboard stop exits 1). The flag waives the exit code and nothing else: same bytes out, and every exit code in a terminal, in a hook and for the gate is unchanged. It is refused by any other command. See the CLI reference.

Works with

Route	Provider	What it reads
minimax-token-plan	MiniMax Token Plan	rolling 5-hour and weekly quota
kimi-code-plan	Kimi Coding Plan	rolling 5-hour and weekly quota
glm-coding-plan	Z.ai (GLM)	percent used and next reset
deepseek-direct	DeepSeek	balance and availability
openrouter-main	OpenRouter	balance and this month's spend

Roadmap, not shipped: Cloudflare Workers AI as a built-in provider (#117), and adding providers from the dashboard (#64).

Also supported:

• Your own OpenAI- or Anthropic-compatible route. Wire compatibility doesn't give you a quota API, so view-limits reads only the documented rate-limit headers from one read-only GET you point it at. Verified with documented-format fixtures, not a live account.
• Xiaomi Token Plan. Experimental and opt-in, macOS only. Set it up with /view-limits:setup-xiaomi — its own command, because it reads Xiaomi's console through your Chrome session rather than a stored key. Two of three live acceptance checks have passed.
• Alibaba Cloud Coding Plan and StepFun Step Plan. Implemented, not live-tested (awaiting owner accounts, #53). StepFun documents no quota API, so that route is shown as unmonitored and never denies.
• GLM. Authentication and error handling are verified live; the quota fields are still checked against fixtures only (#1).

Kimi needs a sk-kimi-* Coding Plan key; a Moonshot sk-* key won't work. Kimi and GLM default to their international hosts, and China accounts set monitor.region: "china" on the route (see Providers). Anthropic isn't a route, because Claude Code surfaces its own state.

Want the details?

• Providers and custom routes: the route catalog, regions for Kimi and GLM, compatible routes, the vl.js provider lifecycle, Alibaba, StepFun.
• Xiaomi Token Plan: requirements, setup, what the numbers mean, fixing access errors.
• Install: the marketplace and checkout installs, and the Codex and OpenCode harnesses.
• Configuration: credential storage, how config.json is assembled, hooks and gate modes.
• Dashboard: the daemon, the loopback page, the phone view, and /view-limits:dashboard.
• The view and its schema: vl.js view, the JSON envelope, what it guarantees.
• CLI reference: every command and exit code.
• Development: architecture, refresh and locking, tests, how releases are cut. Also see credential receipts (the first-run notice), Jev recommendation and Jev readiness (the advisory model recommendation and its readiness gate).

Credits

Built for orchestration in Traycer and similar ADEs, where one orchestrator fans work out across models on different provider accounts. The View-Limits banner is drawn by ascii.rest. The dashboard bundles the Geist fonts (SIL OFL 1.1) and fetches nothing from outside your machine. The Alibaba route reads usage through the official Bailian CLI, and the Xiaomi route reads a console that Xiaomi doesn't document.

License

MIT. See LICENSE.