shunt

shunt adds /shunt:usage to Claude Code. It prints how much headroom the shunt gateway’s shared account pool has left in each rate-limit window, and when each window resets, without opening the gateway’s admin dashboard. The hook answers the command itself and never hands it to the model, so the reading costs no tokens.

The mod is useless without the gateway. shunt, the gateway, is a proxy you put in front of Claude Code through ANTHROPIC_BASE_URL: the model ids you map are diverted to another provider (OpenAI, Codex, xAI, Kimi, DeepSeek and others) while every other model passes through to Anthropic. It can also pool several Anthropic or ChatGPT accounts behind one endpoint, and that pool is what /shunt:usage reports on.

What it does

• Three windows as bars with the percent left and the reset time: 5h (the rolling session window), 7d (the weekly window) and fable (the Fable-scoped weekly window).
• A pool status line (ok, degraded or exhausted), then one row per pooled provider with the same three figures.
• Headroom, not spend. 62% means 62% of the pool’s capacity is unused, averaged over the pool’s active accounts: nine exhausted accounts plus one fresh one read 10%.

Install

Needs Claude Code 2.1.287 or later, a shunt gateway with its usage endpoint switched on, and a client token for it.

claude plugin marketplace add pleaseai/shunt
claude plugin install shunt@shunt

The gateway installs with brew install pleaseai/tap/shunt (or cargo install --git https://github.com/pleaseai/shunt) and listens on http://127.0.0.1:3001 after shunt run. GET /usage is opt-in: add empty [server.auth] and [server.usage] tables to shunt.toml, and export SHUNT_CLIENT_TOKENS="claude-code:<token>" where the gateway runs. Then point Claude Code at it:

export ANTHROPIC_BASE_URL=http://127.0.0.1:3001
export ANTHROPIC_AUTH_TOKEN=<your client token>

Run /shunt:usage. The same marketplace lists seven more plugins (shunt-codex, shunt-xai, shunt-kimi, shunt-deepseek, shunt-zai, shunt-minimax, shunt-mimo). They are ordinary subagent plugins, not mods: each defines subagents on one provider’s models routed through the gateway. None is needed here.

How it works

The mod hooks command.run for shunt:usage only. It reads five environment variables and writes none. The gateway URL is SHUNT_BASE_URL if set, otherwise ANTHROPIC_BASE_URL, with /usage appended; with neither set it asks nothing rather than send your credential to Anthropic’s API. The credential is SHUNT_TOKEN or ANTHROPIC_AUTH_TOKEN as an Authorization: Bearer header, falling back to ANTHROPIC_API_KEY as x-api-key. That is one $.http.fetch per command, to a host the session already talks to unless you set SHUNT_BASE_URL on purpose.

Nothing is stored and nothing reaches the model; a 404, a 401 or 403 or an unreachable gateway each print a plain explanation. Where mods are turned off, the command falls back to commands/usage.md, which asks the model to fetch the endpoint with a tool call and costs a turn.

Limitations

• apiKeyHelper sessions show no credential to the mod, because Claude Code never re-exports the helper’s output. That includes shunt gateway claude, whose launcher scrubs the token variables. Export SHUNT_TOKEN with a [server.auth] client token; the output of shunt gateway token is a different credential and gets a 401.
• An aggregate, not a prediction. Routing also weighs availability, model, session affinity and priority, so a healthy figure does not guarantee your next request is admitted.
• The endpoint reports nothing until a provider is configured with pooled accounts, and never exposes account names, counts or per-account numbers.
• Plugin version 0.1.0 in a pre-1.0 gateway whose config keys may change between releases. The GitHub stars belong to the gateway repository, not to the mod on its own.