🚀 RTK Token Saver

Tool outputs (git diff, grep, find, ls, tree, log dumps...) often eat 30-50% of your prompt budget. RTK detects them and applies smart, lossless compression before the request hits the LLM:

• Filters: git-diff, git-status, grep, find, ls, tree, dedup-log, smart-truncate, read-numbered, search-list
• Auto-detect: No config needed — RTK peeks the first 1KB of each tool_result and picks the right filter.
• Safe by design: If a filter fails, throws, or makes output bigger, RTK silently keeps the original text. Errors never break your request.
• Universal: Works across all formats (OpenAI, Claude, Gemini, Cursor, Kiro, OpenAI Responses) because it runs before any format translation.
• Default ON: Toggle anytime in Dashboard → Endpoint settings.

Without RTK: 47K tokens sent to LLM
With RTK:    28K tokens sent to LLM   (40% saved · same context · same answer)

🧠 Headroom Token Saver

Headroom is optional and runs separately. ExtremeRouter calls Headroom's local /v1/compress endpoint, then keeps normal routing, fallback, auth, and usage tracking:

Client → ExtremeRouter → Headroom /v1/compress → ExtremeRouter → provider

Local setup:

pip install "headroom-ai[proxy]"
headroom proxy --port 8787

Enable in Dashboard → Endpoint → Token Saver → Headroom. Default URL: http://localhost:8787.

Docker examples:

# Headroom service in same Docker network
http://headroom:8787

# Headroom running on host machine
http://host.docker.internal:8787

If Headroom is down or returns an error, ExtremeRouter fails open and sends the original request.

🐴 Ponytail (Lazy Senior Dev)

Ponytail injects a "lazy senior dev" system prompt into every request, biasing the LLM toward minimal, YAGNI-first code — deletion over addition, stdlib over new deps, one-liners over abstractions. Adapted from DietrichGebert/ponytail.

• Lite — Build what's asked, name the lazier alternative.
• Full — YAGNI ladder enforced: stdlib → native → existing deps → one-liner → minimal code.
• Ultra — YAGNI extremist: deletion first, ship the one-liner, challenge the rest of the requirement in the same response.

Without Ponytail: verbose code, extra abstractions, "just in case" scaffolding
With Ponytail:    shortest working diff, no unrequested abstractions, fewer tokens

Never trades away: input validation, error handling that prevents data loss, security, accessibility, or anything explicitly requested. Enable in Dashboard → Endpoint → Ponytail. Stacks with Caveman (output terseness) and RTK (input compression).

🎯 Smart 3-Tier Fallback

Create combos with automatic fallback:

Combo: "my-coding-stack"
  1. cc/claude-opus-4-6        (your subscription)
  2. glm/glm-4.7               (cheap backup, $0.6/1M)
  3. if/kimi-k2-thinking       (free fallback)

→ Auto switches when quota runs out or errors occur

📊 Real-Time Quota Tracking

• Token consumption per provider
• Reset countdown (5-hour, daily, weekly)
• Cost estimation for paid tiers
• Monthly spending reports

🔄 Format Translation

Seamless translation between formats:

• OpenAI ↔ Claude ↔ Gemini ↔ Cursor ↔ Kiro ↔ Vertex ↔ Antigravity ↔ Ollama ↔ OpenAI Responses
• Your CLI tool sends OpenAI format → ExtremeRouter translates → Provider receives native format
• Works with any tool that supports custom OpenAI endpoints

👥 Multi-Account Support

• Add multiple accounts per provider
• Auto round-robin or priority-based routing
• Fallback to next account when one hits quota

🔄 Auto Token Refresh

• OAuth tokens automatically refresh before expiration
• No manual re-authentication needed
• Seamless experience across all providers

🎨 Custom Combos

• Create unlimited model combinations
• Mix subscription, cheap, and free tiers
• Name your combos for easy access
• Share combos across devices with Cloud Sync

🐝 Hierarchical Swarm Engine

A multi-agent orchestration combo strategy for complex, multi-step tasks. Instead of one model answering, a small team of agents collaborates:

• Manager plans the strategy and splits the task
• Staff workers run subtasks in parallel (load-balanced across providers)
• Audit stage reviews worker output
• Manager synthesizes the final answer
• A Smart Gatekeeper triages the prompt first (simple → single-model fast path; complex → swarm)
• Persona-bleed protection keeps each stage's role isolated

Live telemetry streams to the Dashboard → Swarm page (SSE). Configure in Dashboard → Combos → choose the "Hierarchical Swarm" strategy.

🩺 Health Monitor + 🔌 Circuit Breaker

Two reliability layers that keep requests flowing even when a provider degrades:

• Health Monitor records a sliding window of success/failure samples per provider (in-memory ring buffer) and exposes a live SSE feed + the Dashboard → Health page. Spot a provider going red before it ruins your coding session.
• Circuit Breaker wraps every provider in a CLOSED / OPEN / HALF_OPEN state machine. When a provider fails repeatedly, the breaker OPENS and routing auto-skips it (no wasted requests, no timeouts) — then probes with HALF_OPEN to auto-recover the moment it heals.

Both are zero-config; tunable in Settings.

🔐 Per-Key Model Access Control (ACL)

Each API key you mint can carry an allowedModels allow-list. Requests using that key are rejected up-front (403) if they target a model outside the list — so you can hand out scoped keys to teammates, clients, or downstream tools without exposing your full catalog. Managed per-key in Dashboard → Endpoint.

🍪 Cookies Providers

A dedicated Cookies Provider category (above OAuth on the Providers page) for free web-chat services that authenticate via a browser cookie instead of an API key. Paste your cookies (or just the session token) into a multi-line field and ExtremeRouter handles token refresh, WAF cookies, PoW challenges, and SSE translation per site:

• No anti-bot (reliable): ChatGLM, DeepSeek, Qwen, Kimi, Blackbox, T3 Chat, DuckDuckGo, Venice, DouBao, v0, Poe, Copilot, Meta AI (Muse), Adapta, VeoAI
• Anti-bot (best-effort, may be Cloudflare-blocked): Claude Web, ChatGPT Web, Gemini Web
• Plus existing cookie providers: Grok Web, Perplexity Web

Note: Claude/ChatGPT/Gemini web deploy aggressive Cloudflare anti-bot. Plain server-side fetch is often blocked (403) even with valid cookies — these are included best-effort.

🤖 Devin CLI Provider

Devin (by Cognition) is session-based, not OpenAI-compatible. The Devin adapter bridges it: it creates a Devin session, polls for completion, and synthesizes an OpenAI SSE stream back. Models map to Devin agent modes — devin-normal, devin-fast, devin-lite, devin-ultra. Auth is API-key only (cog_...).

📝 Request Logging

• Enable debug mode for full request/response logs
• Track API calls, headers, and payloads
• Troubleshoot integration issues
• Export logs for analysis

💾 Cloud Sync

• Sync providers, combos, and settings across devices
• Automatic background sync
• Secure encrypted storage
• Access your setup from anywhere

Cloud Runtime Notes

• Prefer server-side cloud variables in production:
  • BASE_URL (internal callback URL used by sync scheduler)
  • CLOUD_URL (cloud sync endpoint base)
• NEXT_PUBLIC_BASE_URL and NEXT_PUBLIC_CLOUD_URL are still supported for compatibility/UI, but server runtime now prioritizes BASE_URL/CLOUD_URL.
• Cloud sync requests now use timeout + fail-fast behavior to avoid UI hanging when cloud DNS/network is unavailable.

📊 Usage Analytics

• Track token usage per provider and model
• Cost estimation and spending trends
• Monthly reports and insights
• Optimize your AI spending

💡 IMPORTANT - Understanding Dashboard Costs:

The "cost" displayed in Usage Analytics is for tracking and comparison purposes only. ExtremeRouter itself never charges you anything. You only pay providers directly (if using paid services).

Example: If your dashboard shows "$290 total cost" while using iFlow models, this represents what you would have paid using paid APIs directly. Your actual cost = $0 (iFlow is free unlimited).

Think of it as a "savings tracker" showing how much you're saving by using free models or routing through ExtremeRouter!

🌐 Deploy Anywhere

• 💻 Localhost - Default, works offline
• ☁️ VPS/Cloud - Share across devices
• 🐳 Docker - One-command deployment
• 🚀 Cloudflare Workers - Global edge network

The dashboard tracks your token usage and displays estimated costs as if you were using paid APIs directly. This is not actual billing - it's a reference to show how much you're saving by using free models or existing subscriptions through ExtremeRouter.

Example:

• Dashboard shows: "$290 total cost"
• Reality: You're using iFlow (FREE unlimited)
• Your actual cost: $0.00
• What $290 means: Amount you saved by using free models instead of paid APIs!

The cost display is a "savings tracker" to help you understand your usage patterns and optimization opportunities.

No. ExtremeRouter is free, open-source software that runs on your own computer. It never charges you anything.

You only pay:

• ✅ Subscription providers (Claude Code $20/mo, Codex $20-200/mo) → Pay them directly on their websites
• ✅ Cheap providers (GLM, MiniMax) → Pay them directly, ExtremeRouter just routes your requests
• ❌ ExtremeRouter itself → Never charges anything, ever

ExtremeRouter is a local proxy/router. It doesn't have your credit card, can't send invoices, and has no billing system. It's completely free software.

Yes! The current FREE providers (Kiro, OpenCode Free, Vertex) are genuinely free with no hidden charges.

These are free services offered by those respective companies:

• Kiro AI: Free unlimited Claude 4.5 + GLM-5 + MiniMax via AWS Builder ID / Google / GitHub OAuth
• OpenCode Free: No-auth passthrough proxy, models auto-fetched from opencode.ai/zen/v1/models
• Vertex AI: $300 free credits for new Google Cloud accounts (90 days)

ExtremeRouter just routes your requests to them - there's no "catch" or future billing. They're truly free services, and ExtremeRouter makes them easy to use with fallback support.

Discontinued free tiers (no longer recommended):

• ❌ iFlow: Was free unlimited, now changed to paid (2026)
• ❌ Qwen Code: Free OAuth tier discontinued by Alibaba on 2026-04-15
• ❌ Gemini CLI: Still works, but using it with non-CLI tools (Claude, Codex, Cursor...) may result in account bans — only use if you stick to Gemini CLI itself

Free-First Strategy:

1. Start with 100% free combo:

   hljs language-sql

   Copy

   1. gc/gemini-3-flash (180K/month free from Google)
   2. if/kimi-k2-thinking (unlimited free from iFlow)
   3. qw/qwen3-coder-plus (unlimited free from Qwen)
   

   Cost: $0/month

2. Add cheap backup only if you need it:

   hljs language-bash

   Copy

   4. glm/glm-4.7 ($0.6/1M tokens)
   

   Additional cost: Only pay for what you actually use

3. Use subscription providers last:

   • Only if you already have them
   • ExtremeRouter helps maximize their value through quota tracking

Result: Most users can operate at $0/month using only free tiers!

ExtremeRouter's smart fallback prevents surprise charges:

Scenario: You're on a coding sprint and blow through your quotas

Without ExtremeRouter:

• ❌ Hit rate limit → Work stops → Frustration
• ❌ Or: Accidentally rack up huge API bills

With ExtremeRouter:

• ✅ Subscription hits limit → Auto-fallback to cheap tier
• ✅ Cheap tier gets expensive → Auto-fallback to free tier
• ✅ Never stop coding → Predictable costs

You're in control: Set spending limits per provider in dashboard, and ExtremeRouter respects them.

Claude Code (Pro/Max)

Dashboard → Providers → Connect Claude Code
→ OAuth login → Auto token refresh
→ 5-hour + weekly quota tracking

Models:
  cc/claude-opus-4-7
  cc/claude-opus-4-6
  cc/claude-sonnet-4-6
  cc/claude-haiku-4-5-20251001

Pro Tip: Use Opus for complex tasks, Sonnet for speed. ExtremeRouter tracks quota per model!

OpenAI Codex (Plus/Pro)

Dashboard → Providers → Connect Codex
→ OAuth login (port 1455)
→ 5-hour + weekly reset

Models:
  cx/gpt-5.5
  cx/gpt-5.4
  cx/gpt-5.3-codex
  cx/gpt-5.2-codex

GitHub Copilot

Dashboard → Providers → Connect GitHub
→ OAuth via GitHub
→ Monthly reset (1st of month)

Models:
  gh/gpt-5.4
  gh/claude-opus-4.7
  gh/claude-sonnet-4.6
  gh/gemini-3.1-pro-preview
  gh/grok-code-fast-1

Cursor IDE

Dashboard → Providers → Connect Cursor
→ OAuth login
→ Monthly subscription

Models:
  cu/claude-4.6-opus-max
  cu/claude-4.5-sonnet-thinking
  cu/gpt-5.3-codex

GLM-5.1 / GLM-4.7 (Daily reset, $0.6/1M)

1. Sign up: Zhipu AI
2. Get API key from Coding Plan
3. Dashboard → Add API Key:
   • Provider: glm
   • API Key: your-key

Use: glm/glm-5.1, glm/glm-5, glm/glm-4.7

Pro Tip: Coding Plan offers 3× quota at 1/7 cost! Reset daily 10:00 AM.

MiniMax M2.7 (5h reset, $0.20/1M)

1. Sign up: MiniMax
2. Get API key
3. Dashboard → Add API Key

Use: minimax/MiniMax-M2.7, minimax/MiniMax-M2.5

Pro Tip: Cheapest option for long context (1M tokens)!

Kimi K2.5 ($9/month flat)

1. Subscribe: Moonshot AI
2. Get API key
3. Dashboard → Add API Key

Use: kimi/kimi-k2.5, kimi/kimi-k2.5-thinking

Pro Tip: Fixed $9/month for 10M tokens = $0.90/1M effective cost!

Kiro AI (Claude 4.5 + GLM-5 + MiniMax FREE)

Dashboard → Connect Kiro
→ AWS Builder ID, AWS IAM Identity Center, Google, or GitHub
→ Unlimited usage

Models:
  kr/claude-sonnet-4.5
  kr/claude-haiku-4.5
  kr/glm-5
  kr/MiniMax-M2.5
  kr/qwen3-coder-next
  kr/deepseek-3.2

Pro Tip: Best free option for Claude. No API key, no payment, fully unlimited.

OpenCode Free (No auth, auto-fetch models)

Dashboard → Connect OpenCode Free
→ No login required (passthrough proxy)
→ Models auto-fetched from opencode.ai/zen/v1/models

Pro Tip: Fastest setup. Just connect and start coding.

Vertex AI ($300 free credits for new GCP accounts)

Dashboard → Connect Vertex AI
→ Upload Google Cloud Service Account JSON
→ Enable Vertex AI API in your GCP project

Models:
  vertex/gemini-3.1-pro-preview
  vertex/gemini-3-flash-preview
  vertex/gemini-2.5-flash

Vertex Partner (Anthropic / DeepSeek / GLM / Qwen via Vertex):
  vertex-partner/glm-5-maas
  vertex-partner/deepseek-v3.2-maas
  vertex-partner/qwen3-next-80b-a3b-thinking-maas

Pro Tip: New Google Cloud accounts get $300 credits free for 90 days. Plenty for daily coding.

Example 1: Maximize Subscription → Cheap Backup

Dashboard → Combos → Create New

Name: premium-coding
Models:
  1. cc/claude-opus-4-7 (Subscription primary)
  2. glm/glm-5.1 (Cheap backup, $0.6/1M)
  3. minimax/MiniMax-M2.7 (Cheapest fallback, $0.20/1M)

Use in CLI: premium-coding

Monthly cost example (100M tokens):
  80M via Claude (subscription): $0 extra
  15M via GLM: $9
  5M via MiniMax: $1
  Total: $10 + your subscription

Example 2: Free-Only (Zero Cost)

Name: free-combo
Models:
  1. kr/claude-sonnet-4.5 (Claude 4.5 free unlimited)
  2. kr/glm-5 (GLM-5 free via Kiro)
  3. vertex/gemini-3.1-pro-preview ($300 free credits)

Cost: $0 forever (+ 20-40% token savings via RTK)!

Cursor IDE

Settings → Models → Advanced:
  OpenAI API Base URL: http://localhost:20128/v1
  OpenAI API Key: [from extremerouter dashboard]
  Model: cc/claude-opus-4-7

Or use combo: premium-coding

Claude Code

Edit ~/.claude/config.json:

{
  "anthropic_api_base": "http://localhost:20128/v1",
  "anthropic_api_key": "your-extremerouter-api-key"
}

Codex CLI

export OPENAI_BASE_URL="http://localhost:20128"
export OPENAI_API_KEY="your-extremerouter-api-key"

codex "your prompt"

OpenClaw

Option 1 — Dashboard (recommended):

Dashboard → CLI Tools → OpenClaw → Select Model → Apply

Option 2 — Manual: Edit ~/.openclaw/openclaw.json:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "extremerouter/kr/claude-sonnet-4.5"
      }
    }
  },
  "models": {
    "providers": {
      "@rsalmn/extremerouter": {
        "baseUrl": "http://127.0.0.1:20128/v1",
        "apiKey": "sk_extremerouter",
        "api": "openai-completions",
        "models": [
          {
            "id": "kr/claude-sonnet-4.5",
            "name": "Claude Sonnet 4.5 (Kiro Free)"
          }
        ]
      }
    }
  }
}

Note: OpenClaw only works with local ExtremeRouter. Use 127.0.0.1 instead of localhost to avoid IPv6 resolution issues.

Cline / Continue / RooCode

Provider: OpenAI Compatible
Base URL: http://localhost:20128/v1
API Key: [from dashboard]
Model: cc/claude-opus-4-7

VPS Deployment

# Clone and install
git clone https://github.com/rsalmn/extremerouter.git
cd extremerouter
npm install
npm run build

# Configure
export JWT_SECRET="your-secure-secret-change-this"
export INITIAL_PASSWORD="your-password"
export DATA_DIR="/var/lib/extremerouter"
export PORT="20128"
export HOSTNAME="0.0.0.0"
export NODE_ENV="production"
export NEXT_PUBLIC_BASE_URL="http://localhost:20128"
export NEXT_PUBLIC_CLOUD_URL="https://github.com/rsalmn/extremerouter"
export API_KEY_SECRET="endpoint-proxy-api-key-secret"
export MACHINE_ID_SALT="endpoint-proxy-salt"

# Start
npm run start

# Or use PM2
npm install -g pm2
pm2 start npm --name extremerouter -- start
pm2 save
pm2 startup

Docker

Published images (multi-platform linux/amd64 + linux/arm64):

• Docker Hub: rsalmn/extremerouter
• GHCR: ghcr.io/rsalmn/extremerouter

Quick start (use published image):

docker run -d \
  --name extremerouter \
  -p 20128:20128 \
  -v "$HOME/.extremerouter:/app/data" \
  -e DATA_DIR=/app/data \
  rsalmn/extremerouter:latest

→ Open http://localhost:20128

Build from source (dev):

git clone https://github.com/rsalmn/extremerouter.git
cd extremerouter/app
docker build -t extremerouter .
docker run -d --name extremerouter -p 20128:20128 \
  -v "$HOME/.extremerouter:/app/data" -e DATA_DIR=/app/data extremerouter

Container defaults:

• PORT=20128
• HOSTNAME=0.0.0.0

Useful commands:

docker logs -f extremerouter
docker restart extremerouter
docker stop extremerouter && docker rm extremerouter
docker pull rsalmn/extremerouter:latest   # update to latest

Data persistence: $HOME/.extremerouter/db/data.sqlite on host ↔ /app/data/db/data.sqlite in container.

Environment Variables

Notes:

• Lowercase proxy variables are also supported: http_proxy, https_proxy, all_proxy, no_proxy.
• .env is not baked into Docker image (.dockerignore); inject runtime config with --env-file or -e.
• On Windows, APPDATA can be used for local storage path resolution.
• INSTANCE_NAME appears in older docs/env templates, but is currently not used at runtime.

Runtime Files and Storage

• Main app state: ${DATA_DIR}/db/data.sqlite (SQLite — providers, combos, aliases, keys, settings, usage history)
• Auto backups: ${DATA_DIR}/db/backups/
• Optional request/translator logs: <repo>/logs/... when ENABLE_REQUEST_LOGS=true
• Both ${DATA_DIR} and ~/.extremerouter resolve to the same location in a Docker container — the symlink /root/.extremerouter -> /app/data is created at build time.

Claude Code (cc/) - Pro/Max:

• cc/claude-opus-4-7
• cc/claude-opus-4-6
• cc/claude-sonnet-4-6
• cc/claude-sonnet-4-5-20250929
• cc/claude-haiku-4-5-20251001

Codex (cx/) - Plus/Pro:

• cx/gpt-5.5
• cx/gpt-5.4
• cx/gpt-5.3-codex
• cx/gpt-5.2-codex
• cx/gpt-5.1-codex-max

GitHub Copilot (gh/):

• gh/gpt-5.4
• gh/claude-opus-4.7
• gh/claude-sonnet-4.6
• gh/gemini-3.1-pro-preview
• gh/grok-code-fast-1

Cursor (cu/) - Subscription:

• cu/claude-4.6-opus-max
• cu/claude-4.5-sonnet-thinking
• cu/gpt-5.3-codex
• cu/kimi-k2.5

GLM (glm/) - $0.6/1M:

• glm/glm-5.1
• glm/glm-5
• glm/glm-4.7

MiniMax (minimax/) - $0.2/1M:

• minimax/MiniMax-M2.7
• minimax/MiniMax-M2.5

Kimi (kimi/) - $9/mo flat:

• kimi/kimi-k2.5
• kimi/kimi-k2.5-thinking

Kiro (kr/) - FREE unlimited:

• kr/claude-sonnet-4.5
• kr/claude-haiku-4.5
• kr/glm-5
• kr/MiniMax-M2.5
• kr/qwen3-coder-next
• kr/deepseek-3.2

OpenCode Free (oc/) - FREE no-auth:

• Auto-fetched from opencode.ai/zen/v1/models

Vertex AI (vertex/) - $300 free credits:

• vertex/gemini-3.1-pro-preview
• vertex/gemini-3-flash-preview
• vertex/gemini-2.5-flash
• vertex-partner/glm-5-maas
• vertex-partner/deepseek-v3.2-maas