OpenAI-compatible connector

The OpenAI connector (connectors/types/openai/index.js) exposes an asmltr install as an OpenAI-compatible REST API. Point any OpenAI-style client — the official SDKs, chat UIs like LibreChat, or an OpenRouter-style router — at the install and it's answered by the local Agent SDK through the core, with the same trust + moderation as every other channel.

It serves:

• POST /v1/chat/completions — non-streaming JSON and SSE streaming (stream: true).
• GET /v1/models — advertises a single model id (model_name).
• GET /health — unauthenticated health check.

No provider key — it rides your subscription

There is no execution API key here. The connector runs the local Agent SDK on your Claude subscription. The Bearer key a client sends is asmltr's own inbound auth, not a provider key — it maps the caller to a trust identity so moderation and authorization apply per caller.

---

Authentication

Callers send Authorization: Bearer <key>. Keys map to a trust identity in a gitignored keys.json beside the connector (copy from keys.example.json):

{
  "keys": [
    { "key": "sk-asmltr-CHANGE-ME-long-random-string", "identity": "owner", "username": "openai-client" }
  ]
}

Each identity must be seeded as a principal in the trust store, or the caller is default-denied. Set require_key: false to run open (callers resolve to an anonymous identity) — only sensible behind your own auth.

---

Stateful session model

This is a stateful chat endpoint. The core session holds history and resumes, so a conversation maps to a stable conversation_key:

openai:<instanceId>:<sha1(identity + first-user-message)>

• The first request the connector sees for a key forwards the whole transcript the client sent (so any pre-existing history is included).
• Every request after that forwards only the latest user message — the core session already holds the history, so re-sent history past the latest turn is ignored.

Most OpenAI clients (which append to a message list and re-send the whole thing each turn) work naturally with this: they keep appending, and the connector resumes the same core session, which shows as one card on the dashboard.

Conversation identity

Because the key seeds off the identity plus the first user message, a genuinely new conversation should start with a fresh opening message; reusing an identical first message resumes the earlier session.

Any system messages in the request are passed through as extra system prompt for that turn.

---

Configuration

Discoverable live at GET /types on the manager. From the connector's configSchema:

Field	Default	Purpose
port	3025	HTTP port.
bind_host	127.0.0.1	Bind address.
model_name	asmltr	Advertised model id (what /v1/models returns and responses echo).
require_key	true	Require a Bearer API key.
keys_file	keys.json (beside the connector)	Gitignored key → trust-identity map.

This connector is request/response only — it declares outbound: false, so it is not a POST /send target.

---

Example request

curl -s http://127.0.0.1:3025/v1/chat/completions \
  -H 'Authorization: Bearer sk-asmltr-CHANGE-ME-long-random-string' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "asmltr",
    "messages": [
      { "role": "user", "content": "Summarize the architecture of this project." }
    ]
  }'

Add "stream": true to receive the reply as an OpenAI-style SSE stream (data: {...} chunks terminated by data: [DONE]).

---

Create an instance

curl -s -X POST 127.0.0.1:3024/instances -H 'Content-Type: application/json' -d '{
  "type":"openai",
  "name":"my-openai-api",
  "enabled":true,
  "config":{
    "port":3025,
    "bind_host":"127.0.0.1",
    "model_name":"asmltr",
    "require_key":true
  }
}'

See Connectors for the full manager API.