Never hit a rate-limit again. Run Claude Code, Cursor & 40+ providers through the AI subscription you already pay for — and cut 40–80% of your tokens.
One OpenAI-compatible endpoint for every AI provider. Auto-fallback so you never stop coding · RTK token-saver trims 40–80% off agent loops · Cloak keeps Claude OAuth un-banned · optional P2P mesh so your stack survives anything — even offline. All in one Go binary: no Docker, no Python, no database.
Route Claude · GPT · Gemini · DeepSeek · Groq · Ollama · vLLM through
http://127.0.0.1:2402/v1. Plug it into Claude Code, Cursor, Codex, Cline, OpenClaw, Hermes — anything that speaks OpenAI, Anthropic or Gemini.
A self-hosted alternative to LiteLLM · OpenRouter · 9router — but a single binary, with anti-ban cloaking + a sovereign mesh nobody else ships.
GitHub · AI Agent companion → · Features · Compare · Quick Start · Mesh · FAQ · API
🤖 Works with ANY OpenAI-compatible agent — Claude Code, Cursor, Cline, Codex, Continue, Aider, Hermes, OpenClaw, custom apps. For the deepest integration (thin-body remote-brain, caretaker pipeline, purpose-built subagents) pair it with the recommended companion: github.com/flowork-os/Flowork_Agent.
One brain (this router) + many bodies (any agent) = your full sovereign AI stack.
Modern AI workflows are fragmented. Every CLI, IDE and agent speaks a slightly different API. Every provider bills differently. Your paid subscriptions sit idle while you burn API credits — and a single rate-limit kills your flow.
Flow Router fixes all of it with one local endpoint:
- 🔌 One endpoint, every model. Point any tool at
http://127.0.0.1:2402/v1and reach Claude, GPT, Gemini, DeepSeek, Groq, local models — anything. - 🔑 Use what you already pay for. Drive Claude Code / Cursor through your existing Claude Pro/Max subscription — no extra API key.
- 🥷 Stay un-banned. Claude OAuth requests are cloaked to look like a genuine Claude Code session — a faithful Go port of proven anti-ban logic.
- 🔁 Never stop coding. Priority → round-robin → cost-optimal fallback chains + a 17-rule cooldown/backoff table — one rate-limit just rolls to the next provider.
- 🕸️ Survive anything. Turn on the P2P mesh and routers replicate knowledge host-to-host — leaderless, internet-optional, self-defending.
- 🖥️ Zero ops. One Go binary. No runtime, no DB server. Runs on a Raspberry Pi.
| Flow Router | LiteLLM | OpenRouter | 9router | |
|---|---|---|---|---|
| Deploy | 🟢 single Go binary | Python + Docker | hosted SaaS | Node.js + Next.js |
| No DB / runtime needed | 🟢 | 🔴 | n/a | 🔴 |
| Use your subscription (no API key) | 🟢 Claude/Codex/Copilot/Cursor… | 🟡 keys | 🔴 | 🟢 |
| Claude anti-ban cloaking | 🟢 Cloak | 🔴 | 🔴 | 🟡 |
| Token-saver | 🟢 RTK 40–80% | 🔴 | 🔴 | 🟡 ~20–40% |
| Auto-fallback chains | 🟢 17-rule cooldown | 🟢 | 🟢 | 🟢 |
| P2P mesh / offline-survivable | 🟢 sovereign mesh | 🔴 | 🔴 | 🔴 |
| Shared brain (RAG) | 🟢 FTS5 Memory Palace | 🔴 | 🔴 | 🔴 |
| Runs on a Raspberry Pi | 🟢 | 🟡 | n/a | 🟡 |
Same job as the popular gateways — route every provider through one endpoint — but Flow Router is the only one that's a single binary and ships anti-ban, a token-saver, and a self-defending P2P mesh. Your traffic, your machine, your rules.
| 🔌 Universal endpoint | OpenAI /v1/chat/completions (+ streaming), Anthropic /v1/messages, OpenAI /v1/responses, Gemini /v1beta/models — all served at once |
| 🔄 Full format translation | Transparent OpenAI ⇄ Anthropic ⇄ Gemini conversion via a dual-hop source → openai → target registry — request, response and streaming SSE |
| 🛠️ Tool-calling parity | tool_calls ⇄ tool_use conversion (incl. streaming tool rounds); tool-id sanitisation + empty tool_result stubs prevent the most common Claude 400 |
| 🧩 26 vendor executors | Per-vendor wire-format backends: antigravity · azure · codex · commandcode · cursor (real ConnectRPC protobuf) · gemini-cli · github · grok-web · iflow · jetbrains · kiro · ollama · opencode · perplexity · qoder · qwen · vertex … |
| 📐 Smart params | 22-param OpenAI passthrough, max_tokens auto-bump for tools/thinking, reasoning-content injection, forced-stream collapse, Responses-API event streamer |
| 🔑 Subscription auth | Drive workloads through Claude Pro/Max, Codex, GitHub Copilot, Cursor Pro, Kiro, JetBrains AI, Google Antigravity — no API key |
| 🥷 Claude anti-ban cloaking | Claude OAuth requests cloaked to mirror a real Claude Code session: client tools renamed _cc + 20 native decoy tools, synthetic x-anthropic-billing-header, CC-format fake user_id. Tool names restored in the response. Auto-off for API-key providers |
| 🪪 OAuth & key import | Connect Codex, Cursor, GitLab, iFlow, Kiro, Claude — or paste a token directly |
| 🧭 Live quota fetchers | Pull real upstream rate-limit windows for 13 providers (claude/copilot/codex/gemini/kiro/glm/minimax/qwen/iflow/…) |
| 🔁 Smart fallback | Priority-ordered providers; auto-retry the next on error/rate-limit |
| 🧩 Combos | Group models into one alias with priority / round-robin / random / cost-optimal strategies + per-model combo fallback |
| 💸 Cost-tier routing | Heuristic classifier (char count + code + tool_use + multi-turn) routes simple queries to cheap/local models, honours explicit picks |
| 🪃 17-rule cooldown | Rate-limit / quota / capacity / overloaded text + 401/402/403/404/429/5xx status rules, exponential backoff |
| ✂️ RTK token-saver | 11 auto-detected tool-output compressors (git-diff, grep, ls, tree…) — typical 40–80% token cut in agent loops |
| 🪨 Caveman mode | Appends a "respond tersely" instruction (lite/full/ultra) to save output tokens; code/paths/commands stay exact |
| 🧬 Server-side RAG | FTS5 BM25 cascade over a Memory Palace — any agent that hits the endpoint gets the same retrieved knowledge + skills + persona |
| 🪞 Compounding ingest | Every interaction can be ingested back as FTS-indexed knowledge — all connected agents make the brain smarter together |
| 🛣️ Thin / Pi body mode | FLOWORK_BRAIN_REMOTE lets a light agent body run with no local brain DB — RAG via the router |
| 📊 Usage analytics | Per-day charts, per-provider breakdown, live request stream, cost estimates |
| 🛡️ MITM inspector | Capture, inspect & replay full request/response bodies; local TLS interception with per-SNI cert minting |
| 🚇 Tunnels | Expose securely via Cloudflare Tunnel or Tailscale, with a health watchdog |
| 🌐 Edge proxy deploy | Generate ready-to-ship proxy workers for Cloudflare, Deno Deploy, Vercel |
| 🎬 Media providers | Route embeddings, text-to-image, TTS, STT and web-fetch/search to dedicated backends |
| 🧠 MCP registry | Register Model Context Protocol servers + live tool discovery, behind a spawn allowlist |
| 🔐 Optional login | Password (argon2id) or OIDC, opt-in session enforcement, per-IP login rate limiter |
| 💾 Backups + migrations | Versioned VACUUM INTO snapshots + idempotent schema migrations with auto pre-snapshot |
| 🔒 Secrets at rest | Provider keys + OAuth tokens AES-256-GCM encrypted in SQLite |
| ⌨️ CLI auto-config | Detect & configure 13 popular AI CLIs/extensions in one click |
One OpenAI-compatible endpoint hides the whole pipeline: the router enriches your prompt with knowledge + anti-hallucination antibodies before the LLM, routes to the cheapest capable provider with auto-fallback, and lets a sovereign mesh grow the brain across your fleet.
Flow Router isn't just a single-box gateway. Turn on the mesh and every router becomes a sovereign node in a leaderless, internet-optional, peer-to-peer brain network — designed to keep your AI stack alive even if the cloud, the company, or the internet itself goes dark. No central server. No single point of failure. Your knowledge replicates host-to-host and defends itself from hostile peers.
Router A ◀──── signed packets (ed25519) ────▶ Router B
:2402 :2402
│ mDNS announce (224.0.0.251:5353) │
│ gossip push → 3 random peers / 10s │
▼ ▼
┌─────────────────────────────────────────────────────┐
│ EVERY inbound knowledge packet runs the 9-layer │
│ gauntlet: signature · freshness · karma · quarantine │
│ · PII · injection · near-dup · consensus · promote │
│ pass → promote + reward karma │
│ flag → quarantine reject → drop + penalise │
└─────────────────────────────────────────────────────┘
| Capability | What it does |
|---|---|
| 🪪 ed25519 identity | Each router self-generates a keypair on first boot — its sovereign passport. Private key never leaves the box |
| 📡 Zero-config discovery | Pure-Go mDNS multicast — routers find each other, no seed list, no config |
| ✍️ Signed transport | Every packet ed25519(sha256(...))-signed + dedup'd; tampered/replayed packets rejected at the door |
| 🤝 Gossip propagation | Push-based epidemic broadcast with seen-set dedup; 2-of-3 BFT hook for emergency revocation |
| 🛡️ 9-layer anti-poisoning | Hostile peers can't silently inject knowledge — filter wired into the live receive path, not just a test endpoint |
| ⭐ Karma trust | Peers earn/lose trust; those below the floor are auto-gated out of discovery + gossip; daily decay |
| 🧬 Near-dup detection | Dependency-free trigram-Jaccard — rejects reworded copies, no embedding model, fully offline |
| 🔀 CRDT replication | G-Counter · LWW · G-Set · 2P-Set + vector-clock causal ordering — any merge order converges |
| 🚫 Cloud-metadata firewall | Discovery hard-blocks 169.254.0.0/16 + metadata IPs — mesh can't be tricked into an SSRF pivot |
Honest status: discovery, identity, signed transport, gossip, the 9-layer filter, karma gating, near-dup, CRDT merge, tool-manifest + LoRA-delta validation are implemented, unit-tested, and verified on a running router. WAN bootstrap beyond LAN, and applying a LoRA delta to live weights (needs a fine-tuning runtime this binary doesn't ship), remain on the roadmap. We don't market what we haven't built.
# Build from source (Go 1.25+)
git clone https://github.com/flowork-os/flowork_Router.git
cd flowork_Router
go build -o flow-router-bin .
./flow-router-bin # dashboard + API on http://127.0.0.1:2402Point any tool at it:
Endpoint: http://127.0.0.1:2402/v1
API Key: flr_... (generate in the dashboard, or any string if auth is off)
Connect a provider in 10 seconds: open http://127.0.0.1:2402 → Providers → pick a preset (Claude Pro/Max, OpenAI, Gemini, DeepSeek, Ollama…) → paste key or OAuth login → done.
# Sanity check
curl http://127.0.0.1:2402/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model":"claude-haiku-4-5","messages":[{"role":"user","content":"hello"}]}'| Endpoint | Purpose |
|---|---|
POST /v1/chat/completions |
OpenAI chat (+ streaming) |
POST /v1/messages |
Anthropic native |
POST /v1/responses |
OpenAI Responses API |
GET /v1beta/models · POST /v1beta/models/... |
Gemini-shape |
POST /v1/embeddings · /v1/images · /v1/audio · /v1/search · /v1/web/fetch |
Media + web |
GET /v1/models |
Aggregated model list across providers |
/api/providers · /api/keys · /api/combos · /api/usage · /api/mcp · /api/mesh/* |
Management surface |
Full route surface lives in routes.go.
- Language: Go 1.25 — single static binary, no CGO for core
- Storage: embedded SQLite (
~/.flow_router/db/data.sqlite); optional Memory-Palace brain with FTS5 - Footprint: small binary, low memory — comfortable on a Raspberry Pi
- Quality gate:
go build·go vet·go test·go test -raceall CLEAN before every release; security-audited (11 fixes), runtime-verified 0 panic
Flow Router is the brain. For the matching body — autonomous multi-agent runtime, native FLOWORK_BRAIN_REMOTE thin-mode, full caretaker pipeline (ingestor, training, dashboards) — use:
Works great with any OpenAI-compatible agent (Hermes, OpenClaw, Claude Code, Cursor…); optimal with Flowork_Agent.
Is it free? Yes — Flow Router is free, open-source (MIT), and runs entirely on your machine. No account, no billing, no telemetry. You only pay your own provider/subscription costs (which it helps you not waste).
How is it different from LiteLLM / OpenRouter / 9router? Same core job — one endpoint for every provider — but Flow Router is a single Go binary (no Docker/Python/Node/DB), and it's the only one that ships anti-ban cloaking, an RTK token-saver (40–80%), and a sovereign P2P mesh. OpenRouter is hosted SaaS; LiteLLM needs Python; 9router needs Node. See the comparison ↑.
Will I get banned using my Claude/Codex subscription? Flow Router's Cloak mirrors a genuine Claude Code session (tool renaming, native decoy tools, CC-format headers) to minimise that risk — a faithful port of proven anti-ban logic. It's designed to keep you safe, but no proxy can promise zero risk; use your own judgment and respect each provider's terms.
Does it really run standalone? Yes. One binary, embedded SQLite, no external services. It runs comfortably on a Raspberry Pi. Pair it with Flowork Agent for the full brain+body stack, but the router works on its own.
What's the RTK token-saver? 11 auto-detected compressors for tool output (git-diff, grep, ls, tree…). In agent loops — where tool results dominate the context — it typically cuts 40–80% of input tokens losslessly. Pair with Caveman mode to trim output tokens too.
Does the mesh send my data anywhere? Only if you turn it on. The P2P mesh is opt-in, LAN-first, ed25519-signed, and runs a 9-layer anti-poisoning gauntlet on every inbound packet. Off by default — your keys and traffic never leave the box.
MIT — free to use, modify and self-host.
Flow Router — your AI traffic, your rules, your machine.
⭐ Star this repo if it saves you time or money.
AI gateway · LLM gateway · LLM proxy · LLM router · OpenAI-compatible API · self-hosted · LiteLLM alternative · OpenRouter alternative · 9router alternative · free AI router · token saver · RTK · never hit rate limit · multi-provider · Claude · GPT · Gemini · DeepSeek · Ollama · vLLM · Claude Code · Cursor · Codex · Cline · Hermes · OpenClaw · MCP · Go single binary · Claude anti-ban · Cloak · subscription proxy · P2P mesh · peer-to-peer · decentralized AI · offline AI · CRDT · gossip · ed25519 · anti-poisoning · sovereign AI · RAG · shared brain · Memory Palace · Flowork · 1 brain many bodies