xLyra — Open-source AI gateway
Bring official APIs, subscription accounts and relay stations behind a single endpoint. Your apps talk to one address; xLyra handles routing, protocol conversion and cost estimation.
Change one address,
and you're done
Every upstream you added used to mean another SDK config, another key and another bill. Now your app only needs to know xLyra.
- No longer maintained separately
OpenAI official keyapi.openai.comAnthropic official keyapi.anthropic.comGemini keygoogleapis.comBalances on two relay stationsrelay-*.example.comCodex subscription token refreshOAuthClaude Code subscription authorizationOAuth
client = OpenAI(- base_url="https://api.openai.com/v1",- api_key="sk-…",+ base_url="https://your-xlyra/v1",+ api_key="<downstream API key>",)# Same for the Anthropic SDK; model names stay unchangedOfficial APIs, subscriptions, relays —
added in one click
18 integration types. Authorize Codex and Claude Code subscriptions directly. xLyra refreshes tokens, syncs quotas where the provider supports it, and picks among your accounts.
When an upstream fails,
requests take another road
Every request scores all available upstreams. A failing one goes into cooldown and rejoins the queue once it recovers.
Switching happens before the first byte is written; once a stream has started, the upstream is never swapped.
- 01Health
- 02Latency
- 03Cooldown
- 04Priority
Every attempt stays in the request log: which site, which upstream key, and why it switched.
Speak any protocol — upstreams understand
Chat Completions, Responses, Anthropic Messages and Gemini convert into one another. Pick a downstream protocol:
Downstream · CHAT COMPLETIONSPOST /v1/chat/completions
{
"model": "claude-sonnet-5-5",
"messages": [
{ "role": "system", "content": "You are a helpful assistant" },
{ "role": "user", "content": "Hello" }
],
"stream": true
}Upstream · Claude official · ANTHROPIC MESSAGESPOST https://api.anthropic.com/v1/messages
{
"model": "claude-sonnet-5-5",
"system": "You are a helpful assistant",
"max_tokens": 1024,
"messages": [
{ "role": "user", "content": "Hello" }
],
"stream": true
}One key per app,
every cent accounted for
Upstream credentials stay inside the gateway; apps only get a downstream key. Who can use what, how much, and what it cost — all in one console.
- 1One downstream address that every app points to
- 2Allowlists by model, site or site group
- 3Total, daily and weekly quotas: a key stops when one runs out; daily and weekly quotas reset on their own
- 4Per-key RPM and TPM limits
Every request shows exactly how its cost was calculated
Input, output and cache hits are priced separately from the price table, upstream key multipliers included. Estimated costs are split by key, model and site — no more guessing at month end.
A day with xLyra
Cursor and Claude Code, one address
Some of the team use Cursor, others use Claude Code. Both point at xLyra with their own downstream keys; the gateway decides which upstream and which account serves each request.
The official API got rate-limited. Nobody noticed.
Claude official returned 429, and the request switched to Relay Nova before the first byte. The rate-limited keys cool down, then rejoin the queue.
The support bot ran out of quota
Its key only has a $50 quota. When it is spent that key stops — it never eats into another app's budget. Daily and weekly quotas also reset on their own.
One glance: what each key cost today
Cost trends stacked by model, a ranking by user, cache hit rate. Which app costs the most is obvious at a glance.
And a few more things we thought of
Any machine that runs Docker
will do
One image contains the Go backend and the React console. Your data stays in your own PostgreSQL.
1services:2 postgres:3 image: postgres:17-alpine4 container_name: xlyra-postgres5 restart: unless-stopped6 environment:7 POSTGRES_DB: xlyra8 POSTGRES_USER: xlyra9 POSTGRES_PASSWORD: postgres_password10 volumes:11 - ./postgres:/var/lib/postgresql/data12 networks:13 - xlyra1415 xlyra:16 image: yachiiiiyo/xlyra:latest17 container_name: xlyra18 restart: unless-stopped19 depends_on:20 - postgres21 environment:22 DB_HOST: postgres23 DB_NAME: xlyra24 DB_USER: xlyra25 DB_PASSWORD: postgres_password26 volumes:27 - ./data:/data28 ports:29 - "5801:5801"30 networks:31 - xlyra3233networks:34 xlyra:35 driver: bridge- One imageThe Go backend and the React console ship in the same image.
- First visitOpen the console, create the admin account, then add your first site.
- Before going liveChange the database password and put the console behind HTTPS.
Frequently asked questions
What is xLyra?
xLyra is an open-source AI gateway and control plane. It sits between your applications and AI providers, exposes unified OpenAI- and Anthropic-compatible endpoints, and handles routing, protocol conversion, rate limiting, cost estimation and failover.
Do I need to change my code to use xLyra?
Usually not. Point the base URL of your existing OpenAI or Anthropic SDK at xLyra and use a downstream API key issued in the console.
Which upstreams does xLyra support?
18 integration types are available: OpenAI, Anthropic Claude, Google Gemini, OpenCode Go, DeepSeek, MiniMax, Xiaomi MiMo, Moonshot, Kimi Code, GLM, GLM Code and TypeSafe; Grok connects through xAI device authorization; NewAPI and xLyra can act as cascaded or aggregated sites; Codex OAuth, Antigravity OAuth and Claude Code OAuth accounts are supported as well.
Do requests switch automatically when an upstream fails?
Yes. The routing engine scores every upstream on health, latency, cooldown state and priority, and moves to the next one when a request fails. Switching only happens before the first byte is written; once a stream has started, the upstream is never swapped.
Where is my data stored?
You deploy xLyra yourself. Site credentials, request logs and usage data live in your own PostgreSQL, with S3-compatible backup support.
How do I deploy xLyra?
Start xLyra and PostgreSQL with Docker Compose, then open the console and create the admin account. Before exposing it publicly, change the database password and set up HTTPS, a reverse proxy and access control.
Why not call each provider's API directly?
Calling them directly means maintaining protocols, keys, rate limits, failure handling and billing separately. xLyra gathers all of that in one gateway: your app only needs to know one address, and the control plane does the rest.
What open-source license does xLyra use?
xLyra is released under AGPL-3.0, with the source hosted on GitHub.
Every upstream,
behind one endpoint
Open source, self-hosted, compatible with the SDKs you already use.