SwitchyAI
Menu
Request an audit

HOW IT WORKS

HOW IT WORKS

SwitchyAI sits between your coding agents and the Anthropic and OpenAI APIs. You keep your own accounts, contracts, rate limits and prompt caches, and change one base URL.

REQUEST PATHSELF-HOSTED MODE

CALLER

YOUR AGENTS

IN YOUR NETWORK

SWITCHYAI GATEWAY

YOUR KEYS

ANTHROPIC / OPENAI

ONE BASE-URL CHANGEYOUR ACCOUNTS
THREE STEPSNOTHING CHANGES IN PRODUCTION UNTIL STEP 3

  1. // STEP 1

    CHANGE THE BASE URL.

    Point your agents at the gateway. Your keys and accounts stay yours, and SwitchyAI never resells tokens. The gateway is hosted for small teams and self-hosted in your own cloud for everyone else.

  2. // STEP 2

    SHADOW AUDIT FOR 14 DAYS.

    The gateway runs in dry-run on mirrored traffic, and production is untouched. You get a report of what a once-per-task policy would have saved on your actual workload, net of cache effects, at task-level quality parity.

  3. // STEP 3

    GO LIVE.

    The gateway decides once per task which model and effort level to use, pins that choice, and escalates to the frontier model on failure signals. The ledger and budgets run continuously.

THE LEDGERMETADATA ONLY BY DEFAULT
PER-TASK LEDGER

ONE ROW PER CALL.

Every call through the gateway writes one row to the ledger. Rows hold metadata only by default: prompt bodies are never stored unless you opt in.

// WHAT A ROW CONTAINS

  1. 01TEAM
  2. 02AGENT
  3. 03TASK
  4. 04MODEL SERVED
  5. 05TOKENS, INCLUDING CACHED AND THINKING TOKENS
  6. 06COST
  7. 07WHETHER THE CALL ESCALATED

// AGENT IDENTITY

Each agent is linked to the person or CI pipeline that started it.

// BUDGETS AND KILL SWITCHES

Set a budget per team and per agent, and stop any team or agent with its kill switch.

WHY ONCE PER TASKCACHE-AWARE ON EVERY CALL

// WHY ONCE PER TASK

DECIDE AT THE START.

Prompt caches are per model. Switching models in the middle of a task re-bills the whole context on the new model. So the gateway decides at task start, pins the choice, and only moves when the numbers say the move pays.

// CACHE-AWARE ON EVERY CALL

Cache effects are part of every routing decision, not an afterthought.

// ROUTING ENGINE

Routing uses TypeSafe's Jev decision model today. The engine is swappable.

// LEARNING

One shared routing model plus a private routing model per customer. Both learn from task features and outcomes, never from prompt text or code.

START HERE

REQUEST ASHADOW AUDIT.

Free for 14 days, in dry-run on mirrored traffic. Production is untouched, and you get a written report at the end.

Request an audit