Setting up Triage
Triage is an OpenAI- and Anthropic-compatible endpoint. You send requests to it exactly as you would to either provider; it triages each one, picks a model from your ladder, forwards it through OpenRouter and streams the answer back.
1. Get a key
Open the app and make a workspace. You get a tri_ key once; save it. Then add your OpenRouter key (from openrouter.ai/keys): you pay OpenRouter directly, Triage takes nothing.
2. Point your client at it
Claude Code
export ANTHROPIC_BASE_URL=
export ANTHROPIC_API_KEY=tri_your_key
claudeOpenAI SDK (Python)
from openai import OpenAI
client = OpenAI(base_url="/v1", api_key="tri_your_key")
r = client.chat.completions.create(model="triage", messages=[{"role": "user", "content": "rename userId to accountId"}])
print(r.choices[0].message.content)curl
curl /v1/chat/completions \
-H "Authorization: Bearer tri_your_key" -H "Content-Type: application/json" \
-d '{"model":"triage","messages":[{"role":"user","content":"plan the migration to a sharded setup"}]}' -iThe model name is ignored. Anything goes; triage reads best.
3. Read the receipt
Every response carries these headers:
| Header | What it says |
|---|---|
x-triage-tag | red, orange, yellow or green |
x-triage-score | The severity score, 0 to 1 |
x-triage-tier | Surgeon, Doctor, Nurse or Front desk |
x-triage-model | The model that actually answered |
x-triage-why | The reason, in one line |
How the score is built
Every request starts at 0.15. These rules add or take away points, and the app lists the ones that fired for each request:
| Rule | Points |
|---|---|
| Words that need judgement: debug, design, migrate, refactor, deadlock, race condition, security, trade-offs, edge cases… | +0.18 each, up to +0.60 |
| Words that mean real work: fix, bug, error, test, implement, build, explain, function… | +0.10 each, up to +0.20 |
| Words that mean typing, on a short ask with nothing hard in it: rename, summarise, translate, format, list, extract… | −0.15 each, down to −0.30 |
| A long ask (over 80 words) | up to +0.20 |
| Big context (over ~5k tokens) / heavy context (over ~20k) | +0.10 / +0.20 |
| Code blocks | +0.06 each, up to +0.15 |
| A stack trace | +0.15 |
| Tools to drive (more than 10 tools) | +0.10 (+0.15) |
| Images | +0.08 |
The client asked for deep reasoning (reasoning_effort, thinking) | +0.15 |
| More than 12 turns deep | +0.05 |
Tags: red at 0.70 and above, orange from 0.45, yellow from 0.22, green below that.
Budgets and modes
You set a daily dollar budget for the Surgeon. As it's spent, the score needed to reach the Surgeon rises from 0.70 toward 0.90; once it's gone, critical work goes to the Doctor until midnight UTC.
Strict (default) routes by tag from the first request. Generous sends everything to the Surgeon while more than half its budget is left, then turns strict.
A conversation (recognised by its system prompt and first message) can move up the ladder on a later turn but never down. If a request has tools, rungs whose model doesn't support tools are skipped.
Burning $TRIAGE for credit
In the app, enter an amount and press Burn. The app builds a normal token transfer to the dead address and appends your workspace tag to the calldata (the token ignores it), simulates it, and asks your wallet to sign. After a few confirmations, Triage reads the transaction from the chain, checks the transfer went to the dead address and carries your tag, and credits the workspace with tokens × the token's market price at that moment. Each burn credits once.
House requests are charged at OpenRouter's listed price for the model that answered, with no markup.
Privacy
Triage doesn't store prompts or replies. It keeps, per request: time, tag, score, model, token counts, cost and what was saved. The public board shows the same minus the workspace. Your own OpenRouter key is encrypted at rest with AES-256-GCM.