Support — Quorum

Support

Quorum for macOS. If none of this answers it, write to us — include your macOS version, the Quorum version from the About window, and what you were doing.

Start here. Quorum ships no language model. Until you point it at a model server you control — one running on your own Mac, or a remote endpoint that speaks the OpenAI-compatible chat completions API — there is nothing for an agent to answer with. Almost every first-run problem is that connection.

1. Getting started

Start a model server and note the address its API is served on. Any server that speaks the OpenAI-compatible chat completions API works, on this Mac or remote. The common local ones all do:

LM Studio: Download a chat model, then turn the local server on from the Developer tab. It serves http://localhost:1234/v1 by default, and needs no key. Quorum has a preset for this address.
Ollama: ollama pull llama3.3 then ollama serve. Its OpenAI-compatible endpoint is http://localhost:11434/v1 — note the /v1, which is easy to leave off. No key needed.
llama.cpp: llama-server -m model.gguf --port 8080 serves http://localhost:8080/v1.
vLLM: vllm serve <model> serves http://localhost:8000/v1. If you started it with --api-key, put that key in Quorum too.
A remote provider: Any hosted endpoint with an OpenAI-compatible base URL and the key it issued you. Remote addresses must be HTTPS.

Quorum does not install, download or start any of these, and it does not bundle a model. Whatever you run, and whatever it costs, is between you and whoever publishes it.

Which model? A chat or instruct model, not an embedding or reranking model — Quorum's automatic selection skips those by name, but a server offering nothing else will have nothing to pick. Agents hold a conversation and follow instructions, so a small model will reply but will often ignore a channel's topic; something in the 7B–30B range is where most people find the replies worth reading.

Then, in Quorum: Model servers → Add server, enter the base URL including its API path, add the key if the server needs one, and save. Click Refresh models — the server's models should appear, which is how you know the connection is real. Next, Templates → Set up starter template creates a small team and opens its channel. Type a message and press ⌘↩.

2. The model server will not connect

Saving a server does not test it. Refreshing models does. When a refresh fails:

Check the server is running and serving its API.
Check the base URL includes the API prefix. For most servers that means it ends in /v1, not just a host and port — this is the single most common mistake. See the addresses in Getting started.
A remote address must be HTTPS. Plain HTTP is accepted only for a server on this Mac. That is deliberate and not configurable.
If the server requires a key, check it is entered and still valid.

3. Agents do not reply

Every member of the channel needs a model the server actually reports. A model that was chosen explicitly and has since been unloaded fails the check before the round starts.
A channel marked local-only refuses a remote server by design.
A round that stops earlier than you expected has usually hit the channel's replies per message limit. That is the limit working.
Very large or very slow models can exceed the request timeout. Quorum says so rather than failing quietly.

4. Tools do nothing

Tool access is deliberately narrow. An agent can use a tool only when all three of these are true: the source is attached to that agent, that specific tool is switched on, and the channel has external tools enabled. A tool source that arrived inside a template offers nothing at all until you confirm its address yourself.

5. Agent computers

Agent computers require a Mac with Apple silicon, and are off for every agent until you switch one on.
The guest kernel is downloaded on first use rather than bundled, because it is GPL-licensed and we do not redistribute it. You are shown what will be fetched, from where and how large it is, before anything is downloaded. The archive it arrives in is large.
Commands a model writes run inside that agent's virtual machine and never on macOS. If you were expecting an agent to change a file on your Mac, that is what the files tool and a folder you grant are for, not the computer.

6. Files

Quorum runs in the App Sandbox. Agents reach files only inside a folder you chose through the system's own file panel. A file outside that folder is invisible to every agent — grant a different folder rather than looking for a permission setting.

7. Where your data lives

Your workspace is a JSON file in your Application Support folder that you can open in any text editor, and API keys are in the macOS Keychain. Nothing is stored on a server of ours, because there is not one. The privacy policy sets that out in full.

8. Requirements

macOS 26 or later.
A separate OpenAI-compatible model server, local or remote, that you provide.
Apple silicon, for agent computers only.

9. Contact

Support, bugs and feature requests all go to one address: support@joesapps.net. Refunds for App Store purchases are handled by Apple; write to us and we will point you at the right form.