Run a local model

Run a local model

This one is not an mcp.json paste. Inference runs on hardware you own and your memory still lives in The Cloud, so the setup attaches a machine rather than a client. When it is done, an agent you dispatch runs on your Mac and reads the same workspace every other client reads.

Before you start

• A Cloud account. Create one at thecloud.so. The connection signs in as you, so the account has to exist first.
• A Mac you can leave running. MachineSync is the signed macOS app that attaches a machine as compute. The machine has to be awake for an agent to run on it.
• Disk space for the model weights. An open model is a real download. Pull the smallest one that does your job first.

The one address

There is a single MCP server address. It speaks HTTP JSON-RPC 2.0. You will paste it once:

https://mcp.thecloud.so/mcp

Steps

1. Pair the Mac with MachineSync Download MachineSync from thecloud.so/machinesync, open it on the Mac, and pair it with the one-time code The Cloud shows you. The machine then appears in your Agents view as compute you can dispatch to.
2. Install Ollama and pull a model Install Ollama on that same Mac and pull an open model, for example ollama pull gpt-oss:20b. This is the engine your agents call. The Cloud does not download weights for you.

   ollama pull gpt-oss:20b

3. Pick the local model in the composer In a chat in The Cloud, open the model picker in the composer (Cmd L) and choose a model under the Native group, which is the one tagged Local. That group lists exactly the Ollama tags your paired machine reported it has pulled, so a model you have not pulled will not be offered.
4. Pick which machine runs it If you have paired more than one machine, use the machine picker next to the model picker to choose which one takes the run. With one machine paired, The Cloud picks it for you.

Verify it worked

Start a new chat and send one word: sync. The Cloud answers with its Charter, your memory, and a line confirming the sync. If you get that back, the connection is live and every other tool is available to that chat.

• The composer chip should read your local model's name, not a hosted one, and the machine picker should name your Mac. If the chip falls back to a hosted model, the machine is offline or has not reported that tag yet.

Troubleshooting

• The Native group is empty. The machine has not reported any pulled Ollama tags. Confirm Ollama is running on the paired Mac and that the pull finished, then reopen the picker.
• The model is listed but greyed out. No paired machine currently has that tag. Pull it on the machine you want to run it on, or pick a model that machine already has.
• Runs fall back to a hosted model. The Mac went to sleep or lost its connection. A dispatched run needs the machine awake and paired for the whole run.
• You expected local inference to also make tools local. It does not. Inference runs on your Mac; your workspace, memory, and every tool call still run in The Cloud, which is what keeps a local run and a hosted run reading the same memory.

Next

• Connect into The Cloud for auth, every other client, and the full tool surface.
• Tool reference for every tool and its safety class.
• Safety and consent for what your AI can do without asking, and what it cannot.
• Build on The Cloud to ship a hosted Studio app from the client you just connected.