Like OpenRouter, but for your personal AI subscriptions.

Most people now pay for several AI subscriptions: Claude Pro/Max, ChatGPT Plus/Pro, SuperGrok, GitHub Copilot, Poe, MiniMax, Kimi Code, Z.ai, Alibaba Coding Plan, opencode Go.

Every time one runs out of credits you stop working and start fixing subscriptions: switch models, re-login your harness, repeat.

Subrouter cycles through your subscriptions automatically when one hits a rate limit or runs out of credits. It rotates across accounts of the same provider, and across different providers.

Run only the login commands for subscriptions you have. One subscription is enough to start.

npx @subrouter/cli login anthropic
npx @subrouter/cli login openai
npx @subrouter/cli login xai
npx @subrouter/cli login opencode-go
npx @subrouter/cli login github-copilot
npx @subrouter/cli login poe
npx @subrouter/cli login minimax
npx @subrouter/cli login kimi
npx @subrouter/cli login zai
npx @subrouter/cli login alibaba

Run the same command again to add another account from the same provider. Subrouter rotates through those accounts before moving to the next provider.

If you already logged in through OpenCode, copy those tokens instead of logging in again:

npx @subrouter/cli import opencode

This reads OpenCode auth.json and writes matching accounts into ~/.subrouter/auth.json. Only providers Subrouter supports are imported. Existing accounts with the same identity are updated. The auth file schema is at auth.schema.json.

npx @subrouter/cli status

The built-in default preset ranks the newest model from each provider. When Subrouter resolves a request, it skips providers without an account. You do not need to create a preset.

Choose opencode Go or Pi.

For opencode Go, add @subrouter/opencode to the plugin array in ~/.config/opencode/opencode.json:

{
  "plugin": ["@subrouter/opencode"]
}

Restart opencode Go, then pick the model subrouter/default.

For Pi, install the extension and pick the same model:

pi install npm:@subrouter/pi
pi --model subrouter/default

When Claude hits its usage limit before output starts, Subrouter transparently retries the request with your ChatGPT subscription. When that subscription is exhausted too, it retries with Grok.

Before output starts, quota and authentication failures move through the pool until every subscription is out. Normal request errors return immediately, so Subrouter does not repeat a bad request across every subscription.

The CLI command reference below covers account management, custom presets, cooldowns, and shell completions.

A preset is an ordered list of provider/model candidates. Subrouter skips candidates in cooldown and retries quota or authentication failures on the next account or provider only before output starts. After output starts, tool follow-ups stay on that provider, model, and account until the agent settles. The next user message starts from the preset ranking again.

                opencode Go / Pi
                model: subrouter/default
                       │
                       v
                ┌──────────────────────────┐
                │ subrouter harness plugin │
                │ @subrouter/opencode      │
                │ @subrouter/pi            │
                └──────────┬───────────────┘
                           │
                           v
               for each candidate in order
               (default preset: 1. anthropic
                 2. openai 3. xai 4. opencode-go
                 5. github-copilot 6. poe 7. minimax
                 8. kimi 9. zai 10. alibaba)
                            v
                   inspect each candidate
                            │
      ┬─────────────────────┼─────────────────────┬
      v                     v                     v
  cooldown           quota/auth failure       output starts
                         before output
      v                     v                     v
     skip            cool down account          stream
                            │
                            v
                    try next candidate
                            │
                            v
                 none left? ──> error

Accounts live in ~/.subrouter/auth.json. Log in multiple times to the same provider to build a rotation pool.

Cooldowns are global per machine (~/.subrouter/config.json). Once an account is rate limited, every session and every harness skips it until the cooldown expires.

Presets are ordered lists of provider/model entries. Append #variant to pin reasoning effort, for example openai/gpt-5.5#high. Every preset shows up in opencode Go and Pi as subrouter/<preset-name>.

OpenRouter and CLIProxyAPI expose general-purpose API endpoints. They accept one client protocol and call a provider using another protocol. This requires request conversion, response conversion, and stream parsing inside the proxy.

Subrouter is a transparent subscription router, not a protocol conversion proxy. It chooses the account and model, then lets the harness's existing provider implementation handle the complete request and response protocol.

OpenRouter / CLIProxyAPI
OpenAI request ──> request translator ──> Anthropic / Gemini request
OpenAI stream  <── response translator <── Anthropic / Gemini stream

Subrouter
                    subrouter chooses account + model
                                   v
              ┌────────────────────┴────────────────────┐
              v                                         v
OpenCode request ──> official AI SDK provider    Pi request ──> Pi native provider
              │                                         │
              v                                         v
      provider-native API                      provider-native API

OpenCode already knows how to call providers through the official AI SDK packages. Pi already knows Anthropic Messages, OpenAI Responses, Codex SSE and WebSocket transport, and other provider protocols through its native pi-ai package. Subrouter does not replace either implementation.

This narrow scope keeps Subrouter much smaller and simpler. It does not need a unified message schema or a matrix of protocol translators. That removes common bugs around tool calls, reasoning blocks, media, usage fields, and streaming event order.

Anthropic OAuth only works if requests look like Claude Code CLI requests. The OpenCode adapter and Pi's native provider apply the required identity, tool names, and beta headers.

Run commands directly with npx @subrouter/cli, or install the shorter subrouter command globally:

npm i -g @subrouter/cli

npx @subrouter/cli login [provider] [--input value] # add a subscription to the pool
npx @subrouter/cli import opencode [--from path]    # copy matching OpenCode logins
npx @subrouter/cli logout [provider] [--force]      # remove all accounts for a provider
npx @subrouter/cli account list [--json]     # accounts + cooldown status
npx @subrouter/cli account status [provider] # exits 1 until login completes
npx @subrouter/cli account remove [provider] [n|email] [--force]
npx @subrouter/cli account order --provider anthropic [email protected] [email protected]

account list numbers accounts from 1. Run login again with the same provider to add another account to its rotation pool. In a terminal, login with no provider shows a list to pick from.

import opencode is for machines that already have OpenCode logins. It copies matching tokens from OpenCode auth.json into Subrouter. Use --from only when that file is not in the default OpenCode data directory.

account order sets the fallback order inside one provider. Pass every account email. The first email is tried first.

Tokens live in ~/.subrouter/auth.json. Presets, cooldowns, and live routes live in ~/.subrouter/config.json. Each file includes a $schema URL so editors can autocomplete fields: auth.schema.json and config.schema.json.

npx @subrouter/cli preset create <name> --models 'anthropic/claude-opus-4-6#max,xai/grok-4.6' [--force]
npx @subrouter/cli preset list
npx @subrouter/cli preset show [name]       # includes currently usable candidates
npx @subrouter/cli preset remove [name] [--force]

The order passed to --models is the fallback order. Append #variant to pin reasoning effort for that candidate. preset create checks the variant against models.dev. A session --variant or subrouter/<name>#high still wins over the preset pin. Every preset appears in both harnesses as subrouter/<name>.

npx @subrouter/cli status                   # everything at a glance
npx @subrouter/cli cooldown clear [--force] # retry every account now

Clear cooldowns after you manually upgrade, renew, or reset a subscription plan. Subrouter cannot detect that the provider restored your quota, so cooldown clear removes every provider and account cooldown and lets the next request retry them immediately. Use --force in a non-interactive shell.

Browser and device login runs in the background when an agent or non-interactive shell starts it. Approve the URL, then poll for completion:

npx @subrouter/cli account status anthropic

account status exits 1 while a login is still running, even when older accounts are already stored, so polling it never mistakes yesterday's expired token for today's login.

API-key providers accept --input in non-interactive shells. Destructive commands ask for confirmation in a terminal and require --force elsewhere.

Browser logins finish on a localhost callback server owned by the running login process. Anthropic uses port 53692, OpenAI uses 1455, and Poe prints a random port chosen for that login. When the browser cannot reach the server, copy the final redirect URL and replay it on the machine that started the login:

curl 'http://localhost:53692/callback?code=...&state=...'
# Authentication successful. You can close this window.

The callback endpoint accepts the browser's GET redirect request, so curl can replay the same redirect URL.

subrouter login anthropic
        v
        ├──> callback server listens on 127.0.0.1:53692  <── curl replay works here
        │
        └──> 30 min timeout ──> server closes ──> replay gets connection refused

The window is 30 minutes. After that the callback server closes and curl replay stops working, so run login again. In an interactive terminal with the browser on another machine, set SUBROUTER_MANUAL_OAUTH=1; Subrouter prints the authorize URL, then asks for the final redirect URL.

The default preset is built in. It ranks the newest model from each provider in the order shown above. When Subrouter resolves a request, it skips providers without stored accounts. Create a preset named default to override it.

@subrouter/opencode registers a subrouter provider inside opencode Go via the plugin config hook. Each preset becomes a model. For installation, see Connect your harness.

Pick subrouter/default (or any subrouter/<preset>) as the model. Presets created after opencode Go starts appear on the next opencode Go restart.

Rotate when credits run out. Log in more than one account. When the first subscription hits a rate limit or spends its quota before output starts, Subrouter cools it down and retries the request with the next account. You keep working.

npx @subrouter/cli login anthropic
npx @subrouter/cli login anthropic   # second Claude account
npx @subrouter/cli login openai

Pick subrouter/default as the session model. On a 429 or 402 before output starts, Subrouter retries the request with the next account in the preset.

One model for every agent. Point the session and every agent at a subrouter preset. You log in once per subscription. You do not re-login inside OpenCode when a provider dies. You do not change 100 agent files to swap anthropic/... for openai/....

{
  "model": "subrouter/default",
  "agent": {
    "explore": { "model": "subrouter/default" },
    "plan": { "model": "subrouter/default" }
  }
}

When Claude is out, ChatGPT takes over. When that one is out, Grok takes over. The model id in config stays subrouter/default.

Tasks and subagents. OpenCode primary sessions launch specialized agents through the Task tool. Explore is the usual case: a fast, read-only agent that searches the repo.

Those agents often pin a specific model. When that model hits a rate limit or runs out of credits, the task fails. The parent session then stops with an error, even when other subscriptions still have quota.

Set the agent model to a subrouter preset. Subrouter rotates to the next subscription, so the task continues.

{
  "agent": {
    "explore": {
      "model": "subrouter/default"
    }
  }
}

The same field works in markdown agents under ~/.config/opencode/agents/:

---
description: Fast read-only codebase search
mode: subagent
model: subrouter/default
---

Any other Task subagent works the same way: set model to subrouter/<preset> instead of a single provider model.

                 parent session
                       │
                       v
                    Task tool
                       │
                       v
      explore agent (model: subrouter/default)
                       v
                       ├──> first subscription 429 ──> cooldown
                       │
                       └──> next subscription ──> task continues

@subrouter/pi registers the same presets as models in Pi. It delegates each request to Pi's native provider stream. Poe and Alibaba use Pi's native OpenAI-compatible provider API. Subrouter selects the subscription, but it does not translate requests or responses between provider formats.

Install the extension as shown in Connect your harness. Then use subrouter/default or another subrouter/<preset> model. Presets created after Pi starts appear after /reload or the next restart.

Completions hook into the subrouter command, so this one needs a global install rather than npx:

npm i -g @subrouter/cli
subrouter completions install

Restart your shell (or run autoload -Uz compinit && compinit for zsh). Then Tab works:

subrouter <TAB>          # shows all commands
subrouter pre<TAB>       # completes to "preset"
subrouter login --<TAB>  # shows available options

Completions stay up-to-date automatically. To remove:

subrouter completions uninstall

This pnpm workspace has four packages:

• cli/ — @subrouter/cli: account stores, presets, cooldown state, provider adapters, and the routing engine (RouterModel, an AI SDK LanguageModelV3)
• opencode/ — @subrouter/opencode: the opencode Go plugin plus the provider entry opencode Go loads
• pi/ — @subrouter/pi: a native Pi provider that delegates to Pi's provider streams without format translation
• website/ — private subrouter-website: the Holocron documentation site for subrouter.org

pnpm install
pnpm build
pnpm test

Tests never hit real APIs. Unit tests fake provider endpoints with local HTTP servers.

The e2e tests boot real opencode Go and Pi harness runtimes, point every adapter at local endpoints via SUBROUTER_*_BASE_URL, and assert that a rate-limited provider cycles to the fallback through each complete pipeline. Pi uses in-memory auth, model, settings, and session stores, so tests never read or write the real Pi config.

Set SUBROUTER_MANUAL_OAUTH=1 when the browser that authorizes is not on the machine running subrouter. The localhost callback can never fire there, so browser login switches to asking for the redirect URL instead. Harnesses that drive login remotely (kimaki's Discord /login) set this for you.