routsi — one OpenAI endpoint, every model and agent

One endpoint, many kinds of answerer

◆API models — OpenAI, DeepSeek, OpenRouter… ◆Translated — Anthropic / Gemini ◆Agents — Devin · Codex · Copilot · Claude Code ◆Pull-workers — bring your own ◆Custom Go handlers

See it route

Same endpoint. It picks the answerer.

Every request is POST /v1/chat/completions. What changes is where it goes — and routsi decides.

🪙Easy question → cheap model auto · LOW

auto → cheap

🧠Hard task → real agent auto · HIGH · escalate

auto → agent

🔌Bring your own agent pull-worker

worker

📌Sticky, escalate-only per conversation

sticky

A different job, not a replacement

Gateways swap providers. routsi swaps answerers.

LiteLLM, OpenRouter and Portkey are mature, powerful gateways — hundreds of providers, load-balancing, spend tracking, caching, guardrails, deep observability. If you need those, use them. routsi doesn't replace them; it does one thing they don't: route each request across fundamentally different answerers — API models and full agents — choosing per task.

Capability	routsi	Gateways (LiteLLM / OpenRouter / Portkey)
Route across API models and full agents (Devin/Codex/…)	✓	✗
Pick the answerer per task, automatically	✓	✗
Agent sessions as models (resume Devin/Codex)	✓	✗
Bring-your-own remote worker (outbound-only)	✓	✗
Hundreds of providers/models out of the box	~	✓
Load-balance, retries, fallback across providers	~	✓
Spend tracking, budgets, rate limits, caching	✗	✓
Deep observability / analytics	~ basic	✓
Single self-hosted binary, no external deps	✓	~

✓ built-in · ~ basic / partial · ✗ not a goal. Different tools for different jobs — many people run routsi in front of a gateway.

Who it's for

Built for people who route by more than price.

See exactly what to expect for each profile →

⚡

App builders on the OpenAI SDK

Send easy questions to cheap models and let hard ones escalate — with zero client changes. Just point the base URL at routsi.

model: "auto"

🤖

Teams running coding agents

Devin, Codex, Copilot, Claude Code — addressable behind one OpenAI API, each with a real, resumable session per conversation.

model: "devin/…"

🔒

Self-hosters & privacy-minded

One Go binary you run yourself. Provider keys stay in your env, never in a SaaS path. Auth + mTLS built in.

one binary · no platform

🌐

Distributed / BYO-agent setups

Let a teammate plug their own logged-in agent in as a pull-worker — outbound-only, NAT-friendly, credentials never leave their machine.

routsi worker run

See it live

A dashboard that shows every decision.

Requests, routed-vs-bypass, per-model tokens, latency, escalations, and live worker status — self-contained at /, with /stats (JSON) and /metrics (Prometheus).

routsi live dashboard: request counts, routing split, per-model tokens and latency, and online workers, updating in real time

One endpoint.
Every model and agent.