Harness gateway

OpenAI, Anthropic, Codex, model discovery, streaming, and airplane-mode local workers.

OpenAI-, Anthropic-, and Codex-compatible model calls, model discovery, streaming, and airplane-mode local workers. Use the endpoint returned by GET /v1/teams/{team} — production default https://gateway.s46.dev.

Route inventory

Method	Route	Purpose
`GET`	`/v1/models`	OpenAI-style model list.
`POST`	`/v1/responses`	OpenAI Responses-compatible request.
`POST`	`/v1/chat/completions`	OpenAI Chat Completions-compatible request.
`GET`	`/anthropic/v1/models`	Anthropic-style model list.
`POST`	`/anthropic/v1/messages`	Anthropic Messages-compatible request.
`GET`	`/codex/v1/models`	Codex/OpenAI-style model list.
`POST`	`/codex/v1/responses`	Codex Responses-compatible request.
`GET`	`/v1/workers` · `/v1/workers/{id}`	Airplane-mode local worker discovery.

The gateway also exposes GET /healthz, GET /readyz, and GET /debug/vars — see Operations.

Chat completions

POST/v1/chat/completions→ 200 / SSE

OpenAI Chat Completions-compatible. Omit model to use the team default. Set "stream": true for text/event-stream chunks that terminate with [DONE].

Auth · bearer (production)

curl -X POST https://gateway.s46.dev/v1/chat/completions \
  -H "Authorization: Bearer $S46_ACCESS_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "s46/devstral-small-2-24b",
    "messages": [
      { "role": "user", "content": "Write a short status update." }
    ]
  }'

// Drop-in for the OpenAI SDK — swap the base URL.
import OpenAI from "openai";

const client = new OpenAI({
  apiKey: accessToken,
  baseURL: "https://gateway.s46.dev/v1",
});

const completion = await client.chat.completions.create({
  model: "s46/devstral-small-2-24b",
  messages: [{ role: "user", content: "Write a short status update." }],
});

from openai import OpenAI

client = OpenAI(api_key=access_token,
                base_url="https://gateway.s46.dev/v1")

completion = client.chat.completions.create(
    model="s46/devstral-small-2-24b",
    messages=[{"role": "user", "content": "Write a short status update."}],
)

Responses & Codex

POST/v1/responses

POST/codex/v1/responses

OpenAI Responses-compatible request/response shapes. Codex routes use the same shapes. Streaming requests set "stream": true and receive OpenAI Responses-style events.

curl -X POST https://gateway.s46.dev/v1/responses \
  -H "Authorization: Bearer $S46_ACCESS_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "s46/devstral-small-2-24b",
    "input": "Write a short status update.",
    "stream": false
  }'

const res = await fetch("https://gateway.s46.dev/v1/responses", {
  method: "POST",
  headers: { Authorization: `Bearer ${accessToken}`, "Content-Type": "application/json" },
  body: JSON.stringify({
    model: "s46/devstral-small-2-24b",
    input: "Write a short status update.",
    stream: false,
  }),
}).then((r) => r.json());

res = requests.post(
    "https://gateway.s46.dev/v1/responses",
    headers={"Authorization": f"Bearer {access_token}"},
    json={"model": "s46/devstral-small-2-24b",
          "input": "Write a short status update.", "stream": False},
).json()

Anthropic Messages

POST/anthropic/v1/messages

Anthropic Messages-compatible. The local llama.cpp adapter supports Anthropic tools, tool_choice, tool_use, and tool_result by normalizing them through an internal OpenAI-compatible tool-call representation.

curl -X POST https://gateway.s46.dev/anthropic/v1/messages \
  -H "Authorization: Bearer $S46_ACCESS_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "s46/devstral-small-2-24b",
    "max_tokens": 512,
    "messages": [
      { "role": "user", "content": [ { "type": "text", "text": "Write a short status update." } ] }
    ]
  }'

const msg = await fetch("https://gateway.s46.dev/anthropic/v1/messages", {
  method: "POST",
  headers: { Authorization: `Bearer ${accessToken}`, "Content-Type": "application/json" },
  body: JSON.stringify({
    model: "s46/devstral-small-2-24b",
    max_tokens: 512,
    messages: [{ role: "user", content: [{ type: "text", text: "Write a short status update." }] }],
  }),
}).then((r) => r.json());

msg = requests.post(
    "https://gateway.s46.dev/anthropic/v1/messages",
    headers={"Authorization": f"Bearer {access_token}"},
    json={
        "model": "s46/devstral-small-2-24b",
        "max_tokens": 512,
        "messages": [{"role": "user", "content": [{"type": "text", "text": "Write a short status update."}]}],
    },
).json()

Model discovery

GET/v1/models

GET/anthropic/v1/models

GET/codex/v1/models

GET /v1/models and /codex/v1/models return an OpenAI-shaped list; /anthropic/v1/models returns the Anthropic shape.

curl https://gateway.s46.dev/v1/models \
  -H "Authorization: Bearer $S46_ACCESS_TOKEN"

const { data } = await fetch("https://gateway.s46.dev/v1/models", {
  headers: { Authorization: `Bearer ${accessToken}` },
}).then((r) => r.json());

models = requests.get("https://gateway.s46.dev/v1/models",
                      headers={"Authorization": f"Bearer {access_token}"}).json()

{
  "object": "list",
  "data": [
    {
      "id": "s46/devstral-small-2-24b",
      "object": "model",
      "created": 1893456000,
      "owned_by": "sovereign46",
      "metadata": {
        "displayName": "Devstral Small 2 24B",
        "certified": true,
        "published": true,
        "verificationState": "trusted"
      }
    }
  ]
}

{
  "data": [
    { "id": "s46/devstral-small-2-24b", "type": "model", "display_name": "Devstral Small 2 24B" }
  ],
  "has_more": false
}

Policy & availability failures

Status	Code	Meaning
`403`	`model_not_allowed`	Requested model is not certified/allowed for the caller's team.
`429`	`rate_limited`	Per-IP stream/request rate limit exceeded.
`503`	`model_unavailable`	Model is allowed but no upstream/local serving path is available.

Airplane-mode workers

GET/v1/workers · /v1/workers/{id}

Discover local workers. Current id: local-llamacpp. Worker states: not_configured, unavailable, unhealthy, model_missing, ready. When configured, requests using the s46 model id are rewritten to the backend alias and proxied to llama-server; max tokens are clamped to local config.

curl https://gateway.s46.dev/v1/workers \
  -H "Authorization: Bearer $S46_ACCESS_TOKEN"

const { workers } = await fetch("https://gateway.s46.dev/v1/workers", {
  headers: { Authorization: `Bearer ${accessToken}` },
}).then((r) => r.json());

workers = requests.get("https://gateway.s46.dev/v1/workers",
                       headers={"Authorization": f"Bearer {access_token}"}).json()

{
  "workers": [
    {
      "id": "local-llamacpp",
      "mode": "airplane",
      "kind": "local",
      "protocol": "openai-compatible",
      "state": "ready",
      "models": [
        { "id": "s46/devstral-small-2-24b", "backendModel": "devstral-small-2:24b-instruct-2512-q4_K_M", "state": "ready" }
      ],
      "capabilities": ["chat", "code", "streaming", "tools"]
    }
  ]
}

Upstream proxy behavior

When upstream URLs are configured, the gateway proxies compatible requests without forwarding the caller's Authorization header or cookies. If S46_UPSTREAM_BEARER_TOKEN is set, that token is sent upstream.

Route group	Upstream setting	Prefix behavior
`/v1/*`	`S46_OPENAI_UPSTREAM_URL`	Preserves `/v1/*`.
`/anthropic/*`	`S46_ANTHROPIC_UPSTREAM_URL`	Strips `/anthropic`, after local-model handling.
`/codex/*`	`S46_CODEX_UPSTREAM_URL`	Strips `/codex` before forwarding.

On this page