Local notes and local AI are separate choices. Acres already reads and writes your notes on disk. To have its built-in assistant run on your computer too, connect it to a downloaded model served by Ollama or LM Studio.
This guide covers that built-in assistant. If you want Claude Code, Cursor, Codex or another external client to work in your vault, use the MCP notes guide and check that client’s own model settings.
Choose a runtime your computer can use
Ollama and LM Studio both provide a way to serve models to an app such as Acres. Ollama suits people who already use its local model library. LM Studio provides a desktop interface for downloading, loading and serving models. You only need one to begin.
Check the runtime’s requirements before choosing a large download. LM Studio’s current requirements, for example, support Apple Silicon Macs and explicitly exclude Intel Macs. An Acres build running on your computer does not mean every local AI runtime supports that computer.
Model size, available memory and context length all affect what you can run. Start with a smaller chat model appropriate for your hardware, then try your actual note task. A model which loads but takes too long to answer may not be a useful daily assistant.
For card creation and board actions, choose a model with good tool-use support. Ollama documents function calling; LM Studio distinguishes native and fallback tool-use support. A fluent chat reply does not establish that the model can call the right tool with valid arguments.
Connect Ollama to Acres
Install Ollama from its official quickstart, open it and download a local chat model. Confirm you can get a simple answer in Ollama before troubleshooting the Acres connection. Choose a local model rather than a cloud-hosted option.
In Acres, open Settings → AI, expand Ollama and check the server address. The default for a server on the same computer is:
http://localhost:11434Use Detect to save the address and retrieve the available models. A reachable server with no models is a different state from a server Acres cannot reach. If the list is empty, complete the model download in Ollama and detect again.
Choose the intended local model in Acres’ default-model setting. The default is used when a chat is set to Auto; a chat with an explicit model selection may keep that selection. Check the active chat as well as Settings, especially if you previously used a cloud provider.
You do not need a provider API key for inference with a downloaded model on the local Ollama server. Running Ollama’s cloud models is a different route.
Connect LM Studio to Acres
Install LM Studio, download a suitable chat model and load it. Its server guide explains starting the API server from the Developer tab.
In Settings → AI in Acres, expand LM Studio. For a server on the same computer, the default address is:
http://localhost:1234If you enabled server authentication in LM Studio, enter its matching API token in Acres. Otherwise leave the optional token empty. Use Detect, then select the discovered model as your default or choose it explicitly in the chat.
Acres 0.3.2 added this connection, model discovery and token support. If the LM Studio option is absent, check your installed Acres version against the current release.
A connection error usually needs a server check first: is the server started, is its port correct, and does its token match? A successful model-list response is only the first check. The model still needs to handle a real request.
Verify local inference before relying on it
A localhost address says where Acres sends the request. It does not tell you everything that server does afterward. Ollama also offers cloud models, and a custom server address can point elsewhere.
Use a downloaded local model and inspect the runtime’s active model and route. If you want to disable Ollama cloud features, its official FAQ documents disable_ollama_cloud in its server configuration or OLLAMA_NO_CLOUD=1, followed by a restart. Follow its instructions for your operating system.
Once the runtime and model are ready, disconnect from the network and make a fresh request using harmless sample notes. That checks whether this workflow can operate without internet; it is not an audit of every service or client on your machine. LM Studio’s offline guide explains that downloaded models and the local server can work without connectivity.
Test reading, then one small change
Create a sample board with two short notes whose details you can verify. Start with a request which should leave the source material untouched:
Read the two notes on this board. List the decisions they contain and identify which note supports each one. Leave all cards and the layout unchanged.
Check that the answer uses the actual notes. A generic response may mean the model has answered from its training instead of using the workspace tools. Ask what it retrieved if the source trail is unclear.
Then try a contained write:
Create one new card containing a checklist based on those decisions. Keep the two source cards unchanged and refer to them in the checklist.
Inspect the card and its references. If the model only says it created something, verify that the object exists. Detection, chat and tool actions are three different checks.
In-app rewrites and larger layout changes can be presented as proposals; some additive actions apply immediately. Activity records AI run summaries, while in-app board changes use the app’s undo history. Guarded Activity revert covers supported external agent changes, rather than every built-in AI run. Keep a backup before a larger reorganization and review outputs as you would with a hosted model.
Diagnose the failure you actually have
| Symptom | First thing to check |
|---|---|
| Server not detected | Runtime running, address and port |
| Connected but model list empty | Model download/loading, then Detect again |
| LM Studio rejects authentication | Matching server token |
| Text works, tools fail | Model tool-use support and actual tool calls |
| Very slow or memory errors | Smaller model, shorter context and competing workloads |
| Works online, fails offline | Selected model route and any network-dependent operation |
Keep the first useful workflow narrow: a summary you can trace to two notes or a checklist you can inspect completely. After that, AI mind mapping covers turning material into a larger visual structure.
The Acres license covers the app’s AI connection; local inference does not incur a hosted-model usage bill. It still uses storage, hardware and electricity, and model terms vary. Choose it because the workflow and data route suit your needs, then judge the output on the work it actually does.