Chat & OpenAI-compatible clients

Waired exposes a Local Gateway on http://127.0.0.1:9473 that speaks both the OpenAI and Anthropic wire formats. Any tool that can talk to those APIs — a browser chat UI, an editor extension, a script — can use your own model by pointing at this address. Claude Code and OpenCode are wired up automatically; everything else uses the endpoints below.

Loopback only

The gateway listens on 127.0.0.1 and is never exposed to the network. Requests from other machines reach your model over the encrypted mesh, not this port. Run the client on the same machine as the agent (or tunnel loopback yourself).

Endpoints

API	Endpoint	Base URL for clients
OpenAI chat	POST /v1/chat/completions	http://127.0.0.1:9473/v1
OpenAI models	GET /v1/models	http://127.0.0.1:9473/v1
OpenAI responses	POST /v1/responses	http://127.0.0.1:9473/v1
Anthropic messages	POST /anthropic/v1/messages	http://127.0.0.1:9473/anthropic

Model names

Use one of the Waired aliases as the model. They resolve to whatever you have loaded, so your client config doesn’t change when you switch models:

• waired/default — the default model.
• waired/coding — the coding model.

Other aliases (waired/small, waired/medium, …) and concrete model IDs are listed in the model catalog. Call GET /v1/models to see what your gateway currently advertises.

Authentication

The gateway authenticates requests with a bearer token:

Authorization: Bearer <gateway-token>

Running waired link generates this token and configures the supported coding agents with it. A third-party client needs the same token supplied as its API key. The token is held in your per-user secret store (the system keychain on macOS; a protected file under the agent’s state directory on Linux and Windows).

Turnkey support is still landing

Auto-configured clients today are Claude Code and OpenCode. Pointing other OpenAI-compatible chat UIs (Open WebUI, LibreChat, Jan, …) at the gateway works with the settings above, but supplying the token to them is still a manual step we’re smoothing out. If you hit friction, that’s expected — the endpoints and model names here are stable.

Example request

curl http://127.0.0.1:9473/v1/chat/completions \
  -H "Authorization: Bearer $WAIRED_GATEWAY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "waired/default",
    "messages": [{"role": "user", "content": "Say hi"}]
  }'

For a quick, token-free check from the same machine, use the CLI instead: waired infer "say hi" (see Verify it works).