Skip to the collection

agent-native · LLM Gateway & Routing · Verified 2026-09-26

SageRoute

Listed on The Agent-Native Index

A trajectory-aware model router

What it does

SageRoute is a proxy that changes models based on an agent’s observed execution trajectory instead of classifying only its opening prompt. Every session starts on a cheap model. The router reconstructs tool calls and results from each request, detects loops, repeated error classes, rewrite/retest thrashing, and lack of successful execution, then asks the Levanto Sage decision API whether to continue, switch models, restart with cleaned context, or escalate to a human.

The project is explicitly early. It has no published release/tag, stores session state in one process, and is not battle-tested at scale. The README documents a real cheap-to-strong run but not a multi-task benchmark, aggregate savings study, or live proof for every ladder action. Live decision quality also depends on the external Levanto Sage API; an offline deterministic stub exists for development, and live Sage failures default to continuing rather than taking the agent down.


Agent-facing interfaces

How agents onboard

Interaction pattern: OpenAI/Anthropic-compatible proxy

git clone https://github.com/codejunkie99/sageroute.git
cd sageroute
bun install
echo "SAGE_API_KEY=lv_..." > .env
bun run serve                     # http://127.0.0.1:8787

export OPENAI_BASE_URL=http://127.0.0.1:8787/v1
export OPENAI_MODEL=sageroute

See the full onboarding section in the dossier.

Protocol surface

Interface Detail
POST /v1/responses Native routed path for Codex and OpenAI-compatible clients
POST /v1/messages Anthropic Messages adapter for Claude Code, translated into the same evidence model
POST /v1/chat/completions Passthrough path for non-routed aliases
GET /v1/models Router alias plus configured provider/model names
GET /v1/sageroute/sessions Per-session tier, cost ledger, switches, and decision history
GET /health Liveness endpoint
CLI serve, init, check, provider selection, and OAuth login/import/status commands

Verification

Last verified 2026-09-26.

Sources: