LINA Router

Never stop coding. Route any AI tool to 40+ providers — and let the cheap ones write the boring parts.

Get Started → · Docs · Site

| 40+ | 100+ | 20–40% | $0 | |:---:|:---:|:---:|:---:| | providers | models | tokens saved | to start |

What is this?

LINA Router is a local proxy that sits between your AI coding tools (Claude Code, Cursor, Cline, Codex, anything OpenAI-compatible) and the AI providers themselves. It does three things you didn't know you needed:

1. Compresses tool outputs before they reach the model. git diff going from 4,200 tokens to 2,600. Every request, automatically.
2. Falls back gracefully when your subscription quota runs out — to cheap APIs, then to free models. You never see a rate limit.
3. Speaks every dialect. OpenAI, Anthropic, Gemini, custom — point your tool at localhost:20128/v1 and it just works.

Built for engineers who hate paying $50/month for tokens they could have saved. Runs entirely on your machine.

📐 How it routes

graph LR
    CLI[Your CLI Tool<br/>Claude Code · Cursor · Cline · Codex]
    LINA[LINA Router<br/>localhost:20128]
    Sub[Tier 1<br/>Subscription<br/>Claude · Copilot · Codex]
    Cheap[Tier 2<br/>Cheap APIs<br/>GLM · MiniMax · DeepSeek]
    Free[Tier 3<br/>Free Forever<br/>Kiro · OpenCode · Vertex]

    CLI -->|OpenAI-compatible| LINA
    LINA --> Sub
    LINA --> Cheap
    LINA --> Free

    style LINA fill:#D97757,stroke:#A85638,color:#fff
    style Sub fill:#FAF9F5,stroke:#D97757
    style Cheap fill:#FAF9F5,stroke:#D97757
    style Free fill:#FAF9F5,stroke:#D97757

⚡ Quick Start

npm install -g lina-router
lina-router

Dashboard opens at http://localhost:20128. Connect a free provider (Kiro takes 10 seconds, no signup), then point your tool:

export ANTHROPIC_BASE_URL=http://localhost:20128/v1
export ANTHROPIC_API_KEY=<your-lina-router-key>
claude

That's the entire install.

🧭 How fallback decides

When a request comes in, LINA Router asks three questions in order. The first yes wins:

flowchart TD
    Req[Incoming request]
    Q1{Subscription quota<br/>remaining?}
    Q2{Cheap-tier budget<br/>under limit?}
    Q3{Free provider<br/>available?}
    Sub[→ Route to subscription<br/>Claude / Copilot / Codex]
    Cheap[→ Route to cheap API<br/>$0.20–0.60 per 1M tokens]
    Free[→ Route to free<br/>Kiro · OpenCode · Vertex]
    Fail[Return error]

    Req --> Q1
    Q1 -->|yes| Sub
    Q1 -->|no| Q2
    Q2 -->|yes| Cheap
    Q2 -->|no| Q3
    Q3 -->|yes| Free
    Q3 -->|no| Fail

    style Req fill:#FAF9F5,stroke:#141413
    style Sub fill:#D97757,stroke:#A85638,color:#fff
    style Cheap fill:#D97757,stroke:#A85638,color:#fff
    style Free fill:#D97757,stroke:#A85638,color:#fff
    style Fail fill:#fff,stroke:#A85638

You can reorder tiers, lock specific models, or run multi-account round-robin within a tier. Whatever keeps you coding.

💎 Features

| | | |---|---| | 🚀 RTK Token Saver | Auto-compresses git diff, grep, ls, tree tool outputs before they reach the model. 20–40% smaller, same fidelity. | | 🔄 Auto Fallback | Subscription → Cheap → Free. Zero downtime, zero intervention. You'll know it switched because the dashboard told you. | | 👥 Multi-Account | Got three Claude Code accounts? LINA round-robins between them to dodge rate limits and squeeze every quota. | | 🌐 Universal | One endpoint. Works with Claude Code, Cursor, Cline, Codex, Copilot, Continue, OpenCode — anything OpenAI-compatible. |

🪶 RTK — the boring genius

Most of your tokens aren't prose. They're tool outputs — git diff against a 50-file refactor, grep -r across a monorepo, ls -la of a node_modules tree. RTK looks at those, strips the redundant parts, and forwards a leaner version. The model still gets what it needs. Your wallet doesn't.

flowchart LR
    Tool[CLI tool output<br/>git diff · grep · ls]
    Detect{Detected as<br/>tool_result?}
    Compress[RTK Compress<br/>strip redundant lines<br/>preserve semantics]
    Forward[Forward to model<br/>20–40% smaller]
    Forward2[Forward as-is]

    Tool --> Detect
    Detect -->|yes| Compress
    Detect -->|no| Forward2
    Compress --> Forward

    style Tool fill:#FAF9F5,stroke:#141413
    style Compress fill:#D97757,stroke:#A85638,color:#fff
    style Forward fill:#D97757,stroke:#A85638,color:#fff
    style Forward2 fill:#FAF9F5,stroke:#141413

One request rarely matters. Ten thousand do.

🛠️ Tools that work today

🌐 Providers

🔐 OAuth (your subscription)

🆓 Free forever

🔑 API key (40+ more)

OpenRouter · GLM · Kimi · MiniMax · OpenAI · Anthropic · Gemini · DeepSeek · Groq · xAI · Mistral · Perplexity · Together · Fireworks · Cerebras · Cohere · NVIDIA · SiliconFlow · Nebius · Chutes · Hyperbolic · and 20 more, plus any OpenAI/Anthropic-compatible endpoint.

💰 Pricing reality check

| Tier | Provider | Cost | Best for | |---|---|---|---| | RTK (built-in) | LINA Router | $0 | Saves 20–40% on every request | | Subscription | Claude Code (Pro/Max) | $20–200/mo | Already a subscriber | | Subscription | Codex (Plus/Pro) | $20–200/mo | OpenAI heavy users | | Subscription | Copilot | $10–19/mo | GitHub-native | | Cheap | GLM-5 | $0.6 / 1M tokens | Budget backup | | Cheap | MiniMax M2.7 | $0.2 / 1M tokens | Cheapest paid option | | Cheap | Kimi K2.5 | $9 / mo flat | Predictable monthly | | Free | Kiro AI | $0 | Unlimited Claude 4.5 | | Free | OpenCode Free | $0 | No auth, no signup | | Free | Vertex AI | $300 credits | New GCP accounts |

The combo: RTK + Kiro + OpenCode = $0 cost, −40% tokens. We were not paid to recommend it; it's just the math.

No, LINA Router never bills you. The dashboard shows "saved $290 this month" — that's the money you didn't spend. It's an open-source binary running on your laptop. There is no card to charge.

🎯 Who this is for

"I have Claude Pro and keep hitting limits." LINA round-robins across your accounts, then falls back to Kiro free when you're out. Same model, no interruption.

"My GLM API bill keeps growing." RTK alone cuts 30% off every request. Free Kiro tier absorbs the rest. Bill goes down.

"I want to try Cursor without paying." Point Cursor at localhost:20128/v1. Connect Kiro. You now have Cursor running on free Claude 4.5.

📦 Install variants

git clone https://github.com/spooky-may/lina-routerrr.git
cd lina-routerrr
pnpm install
pnpm dev

Dashboard at localhost:20128. Hot reload enabled.

docker build -t lina-router .
docker run -d \
  --name lina-router \
  -p 20128:20128 \
  -v lina-router-data:/app/data \
  lina-router

Data persists in the lina-router-data volume.

git clone https://github.com/spooky-may/lina-routerrr.git
cd lina-routerrr && pnpm install && pnpm build
pm2 start ecosystem.config.cjs
pm2 save && pm2 startup

🧠 Tech under the hood

• Runtime: Node.js 18+, Next.js 16 (App Router)
• DB: SQLite (better-sqlite3) for local persistence
• Cloud worker: Cloudflare Workers + D1 (optional, for cross-device sync)
• Tunnel: Tailscale Funnel + Cloudflare Quick Tunnels
• MITM (for Copilot/Kiro): mitmproxy with custom Node bridge
• OAuth: 14+ provider-specific implementations (Claude, GitHub, Gemini, Cursor, …)
• Telemetry: in-process event ring buffer + JSONL drain
• Health: per-provider circuit breaker (closed → open → half-open)

• All routing logic lives in open-sse/ (handlers, account fallback, combo routing)
• Dashboard is a Next.js app at /dashboard
• Provider OAuth flows in src/lib/oauth/services/{provider}.js
• Database repos in src/lib/db/repos/
• MCP plugin bridge: src/lib/mcp/stdioSseBridge.js exposes stdio MCP servers over SSE
• Observability: src/lib/analytics/eventTelemetry.js (events) + src/lib/healthcheck/healthMonitor.js (circuit breakers)

LINA Router exposes an OpenAI-compatible /v1/chat/completions endpoint at localhost:20128. Models are referenced via prefixes that map to providers:

| Prefix | Meaning | |---|---| | kr/ | Kiro AI | | cx/ | Codex / OpenAI | | ag/ | Antigravity | | gh/ | GitHub Copilot (MITM) | | or/ | OpenRouter | | (any model id) | Default routing rules apply |

Example: kr/claude-sonnet-4.5 routes through Kiro using Claude 4.5.

Full OpenAPI spec at /api/openapi.json once you're running locally.

• Dashboard won't open? Check port 20128. lsof -i :20128 (mac/linux) or netstat -ano | findstr :20128 (windows).
• MITM error for Copilot/Kiro? Trust the LINA root CA. Dashboard → Settings → Install Certificate.
• OAuth callback fails? Make sure no other tool is bound to localhost:20129. That's the callback port.
• High RAM usage? SQLite cache. Run Settings → Maintenance → Compact DB.

❓ FAQ

Is LINA Router free? Yes. MIT licensed. No paywall, no telemetry-to-server, no "premium tier." You pay providers directly; LINA just routes.

Does it work offline? The router does (after install). Providers, obviously, don't.

Will my OAuth tokens leak? Tokens live in ~/.lina-router/ (or DATA_DIR) as SQLite rows. Same security model as your shell history. No cloud sync unless you explicitly enable it.

Can I add a custom provider? Yes. Dashboard → Providers → Custom → paste the endpoint. Anything OpenAI-compatible works out of the box.

Why "LINA"? It's just a name. No, there's no token.

🙏 Built on

• RTK — the token compressor that does the heavy lifting
• Next.js, Tailwind, better-sqlite3
• Cloudflared, Tailscale

LINA Router · GitHub · Site · Docs

MIT License — free as in beer, free as in speech.