๐—๐’๐ˆ

Understand XSI Router in 3 Minutes: Architecture and Quick Start

Posted October 5, 2025ย โ€ย 3ย min read

You probably care about two things: 1) what this is, 2) can I use it now. This post first shows how to run it in 60 seconds, then explains the routing and security design in simple words.

What it is โ€” one sentence

An enterpriseโ€‘grade AI API entry. Change your base URL and use your own provider keys (BYOK) to call OpenAI / Anthropic / OpenAI Codex / Claude Code โ€” with builtโ€‘in rate limits, allowlists, smart routing, and auditable usage.

What it supports (out of the box)

  • OpenAI compatible: /v1/chat/completions, /v1/responses, /v1/embeddings, /v1/audio/*, etc.
  • Anthropic compatible: /v1/messages
  • OpenAI Codex / gptโ€‘*-codex for code completion
  • Claude Code (IDE coding assistant): see โ€œClaude Codeโ€
  • Others: Gemini, Mistral, Cohere, DeepSeek, Grok, Perplexityโ€ฆ (plug in your provider keys)

60โ€‘second quick start

  1. Get an API key (your provider key)

  2. Just change the base URL:

export XAI_API_KEY="sk-Xvs..."
curl https://api.xsirouter.com/v1/chat/completions \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5-nano","messages":[{"role":"user","content":"hello"}]}'

Node.js:

import OpenAI from "openai";
const client = new OpenAI({ apiKey: process.env.XAI_API_KEY, baseURL: "https://api.xsirouter.com/v1" });
const res = await client.chat.completions.create({ model: "gpt-5-nano", messages: [{ role: "user", content: "hello" }] });

Python:

from openai import OpenAI
client = OpenAI(api_key="sk-Xvs...", base_url="https://api.xsirouter.com/v1")
resp = client.chat.completions.create(model="gpt-5-nano", messages=[{"role":"user","content":"hello"}])

No SDK rewrites โ€” just change the base URL.

Architecture (plainโ€‘English)

We talk about three planes. The names sound academic; the ideas are simple:

  1. Control plane (rules)
  • Model Mapper โ€” alias old names to new models; switch without app changes.
  • Level Mapper โ€” send different models to different pools by cost/latency/stability.
  • Switch Over โ€” explicit โ€œprimaryโ†’backupโ€ relations (e.g., 1โ†’2) with clear triggers.
  • Resources allowlist โ€” restrict which API paths are callable.
  • Model Limits โ€” hard perโ€‘model RPM/TPM caps to prevent cost spikes.
  1. Routing plane (fast path)
  • Sameโ€‘level roundโ€‘robin with atomic indices and fineโ€‘grained locks.
  • Sleep on errors โ€” bad keys cool down without punishing healthy keys.
  • Explicit switch over โ€” only after sameโ€‘level retries fail; no blackโ€‘box fallbacks.
  • Transport tuning โ€” hot pools, HTTP/2, large buffers, streaming with backpressure.
  1. Usage plane (see clearly)
  • Perโ€‘user/model metering with daily/monthly rollups (requests/tokens/cost).
  • Governance in one place: IP/Model/Resource allowlists + dashboards/logs.

Security & compliance (four lines)

  • BYOKโ€‘first: always your keys; no silent fallbacks; clear billing/audit boundaries.
  • Key encryption: sanitize + encrypt at rest; decrypt justโ€‘inโ€‘time; never log secrets.
  • Allowlists enforced: IP / Models / Resources with inheritance that never exceeds the parent.
  • Fail fast: exceed limits or violate policy โ†’ explicit failure, not invisible overspend.

Cost & governance (straightforward)

  • Direct billing to providers; your discounts apply.
  • Hard caps per user/model; protect first, then serve.
  • Zero markup in selfโ€‘hosted mode; transparent policies, predictable totals.

Next steps