LocalASOMCP and AI agents
Local AI for ASO with Ollama: keep keyword work on your Mac
Ollama lets LocalASO run its AI steps on a model on your own Mac. The docs describe it as free and fully offline, with quality that depends on the model you choose, so it suits people who do not want keyword drafts sent to a cloud model.
Updated
What the docs say
In LocalASO's AI provider list, Ollama needs Ollama running locally at http://localhost:11434 and a model pulled, for example ollama pull llama3.1. The cost is free. The notes are short: quality depends on the model, and you should use a strong model for localization. You choose the provider in Settings under AI provider, and the Test button checks that it answers. The full table is in the AI providers docs.
Three data paths, kept separate
Privacy claims go wrong when they blur different data flows. LocalASO has three:
| Data | Where it lives or goes |
|---|---|
| Boards, keywords, ranks, listings | A database on your Mac |
| AI prompts and answers | To the provider you picked. With Ollama, to the local Ollama process at localhost |
| Rank checks | Public Apple search requests, which need an internet connection |
AI calls go from your Mac straight to the provider you picked, and LocalASO itself has no server. With Ollama that provider is on the same machine. The rank checks are separate: they are public Apple search requests, and they are the one thing an agent session through the MCP server sends out, as described in the MCP docs. So "everything stays local" is accurate for your data and the AI step, but rank checks still need the network.
Choosing a model
The docs do not benchmark models, and neither does this page. They say quality depends on the model and that localization deserves a strong one. A practical test costs a few minutes: pick five keywords you know well in a language you read, localize them with a small model and with a larger one, and compare the phrases by eye and by rank data. Localizations are cached by keyword and country, so retesting a different model means clearing or choosing new terms.
Ollama and the MCP server
If an agent such as Claude Code calls board_add_keywords with mode: "localize", the server needs a provider that works without a stored key. The MCP docs say the default Claude Code CLI does and do not mention Ollama in that context, so test Ollama through the MCP route before relying on it; the safe route is to localize inside the app when Ollama is your provider. Then use the agent for reading and checking: board_get and keywords_check do not need an AI provider at all. Note too that the agent client itself, for example Claude Code, may use a cloud model for its own reasoning; the tool results you let it read pass through that model.
When Ollama is not the best fit
- Small models can produce unnatural phrases in less common languages. Check results, as described in localize App Store keywords with AI.
- If you want the strongest listing rewrites, the docs' smart-model setting applies to the Claude Code CLI and Anthropic API only, so Ollama uses whatever model you pulled.
Next step
Install Ollama, pull a model, select it in LocalASO's Settings and press Test. Then read the keyword research guide to decide which keywords are worth localizing, see AI agent prompts for what to ask, and get the app from the download page.
Questions
- Is everything offline with Ollama?
- The AI step is. Rank checks still call the public Apple search endpoint, so you need an internet connection for those.
- Which Ollama model should I use?
- The docs use llama3.1 as an example and advise a strong model for localization, because quality depends on the model.
Do this for every country, in minutes