Skip to main content

Connect your software

For Claude Desktop, Claude Code, OpenCode or Codex, start with Aventail Connect. Open Connected apps in Link and let it configure the supported app for you.

For your own software or another client, follow the API walkthrough below. Keep Link running with a serving local model or an applied cloud route, and run the first test on the same machine as Link.

AVENTAIL / INTEGRATION

Send a local request

Full view

Model Serving in Control

Open Serving, then expand API Usage to find the local connection examples.

Slide 1 of 4: Model Serving in Control.

Use the exact model ID and your actual serving port. Use the arrows or progress marks to move between slides.
AVENTAIL / INTEGRATION

Send a local request

Full view

Model Serving in Control

Open Serving, then expand API Usage to find the local connection examples.

Slide 1 of 4: Model Serving in Control.

Read the API steps and copy commands

Send a request to the model running through Link, then use the same endpoint in your application.

Before you start​

Complete setup with local serving or a supported applied route. Keep Link running and open Terminal on the same Mac. Aventail setup serves on port 8100. If you configured another endpoint, use its actual port.

Tested: Aventail Link 2.0.0 · Llama 3.2 3B Instruct Q4_K_M · Apple Silicon · 2 October 2026.

Cloud-first routing​

For Link 2.4.0, first check Routing and backups. A test through /v1/chat/completions on 13 November returned a successful Anthropic Haiku response, with Control reporting location: anthropic. A second request naming the local backup still followed the configured cloud-first route. The requested model name and a localhost address do not prove local execution. Inspect the applied route and request records.

A successful cloud response does not test automatic failover. Keep provider keys out of client examples; use the saved, device-authorised connection in Control.

1. Find your model ID​

In Link, open Models, select the model, expand Model details and use Copy beside Model ID (called Model identifier in 2.3.2). Copy the identifier rather than typing the model’s display name: the two can differ. You can also copy Link’s address under Connected apps → Connect another app yourself, with Copy address.

Alternatively, query the local API. Replace 8100 if you chose a different port:

curl --fail-with-body --silent --show-error \
http://127.0.0.1:8100/v1/models

Copy an id from the returned data array. Use that exact value, including spaces and punctuation, in the next request.

Local connection

Use 127.0.0.1 for requests on this Mac, not the bind address 0.0.0.0. Serving can listen on all interfaces; review network exposure before connecting other machines.

2. Send a request​

Save this as request.json, replacing the model ID with yours:

{
"model": "Llama-3.2-3B-Instruct-Q4_K_M",
"messages": [{"role": "user", "content": "Reply with exactly: Aventail is ready."}],
"max_tokens": 32,
"temperature": 0,
"stream": false
}

Run:

curl --fail-with-body --silent --show-error --max-time 180 \
-H 'Content-Type: application/json' \
--data-binary @request.json \
http://127.0.0.1:8100/v1/chat/completions

3. Check the response​

A successful test needs a successful HTTP response and non-empty choices[0].message.content. The recorded response included:

{
"choices": [
{
"finish_reason": "stop",
"index": 0,
"message": {
"role": "assistant",
"content": "Aventail is ready."
}
}
]
}

The wording can vary. The response may report the model file path instead of the request's alias.

Connect an application​

Use http://127.0.0.1:8100/v1 as the base URL for an OpenAI-compatible client on this Mac, with the model ID returned above. Browser clients also need an allowed origin; see the Link integration reference.

For app-specific setup, see Continue in VS Code, Open WebUI, Obsidian, or OpenClaw. You can also chat in Workspace.

If the request fails​

SymptomNext action
Connection refusedConfirm Link is running, local serving or a supported route is available, and the port matches the copied Link address.
Model-name/router errorCall /v1/models again and copy an exact id.
HTTP success but empty contentInspect finish_reason. If it is length, increase max_tokens for the model and task.
First request is slowAllow time for engine/model startup. Retain any error text rather than repeatedly redeploying the model.

To stop serving, use Stop serving in Link, or the model’s control for the intended device in Control. Finish active requests first. The model stays installed.