# Run WireCat on local models (/en/docs/desktop/local-models)

WireCat keeps your messages, mail, notes and files on your computer. Some features also send text, pictures or a
search question to an AI model, and so does the agent you connect to WireCat. When that model is a cloud service, your
private text goes to that company. This guide shows how to run a model on your own computer and point every feature
and your agent at it, step by step, so that what they read stays with you. You need a computer that can run a model,
ideally with a graphics card. A few parts of WireCat cannot use a local model yet, and the end of the page lists them.

## 1. Run a model on your computer [#1-run-a-model-on-your-computer]

[Ollama](https://docs.ollama.com/quickstart) is the easiest start, and WireCat Desktop knows it by name.

1. Install Ollama and download a model with `ollama pull <model>`. For reading pictures, pick a model that can see
   them, such as `llava`.
2. Ollama listens on `127.0.0.1:11434`, which WireCat expects. Keep that address.
3. Turn off Ollama's cloud models. A model whose name ends in `cloud` runs on Ollama's servers, and its web search goes
   out too. [Ollama's FAQ](https://docs.ollama.com/faq#how-do-i-disable-ollamas-cloud-features) describes local only
   mode: set `"disable_ollama_cloud": true` in `~/.ollama/server.json`, or `OLLAMA_NO_CLOUD=1`, then restart Ollama.
4. If you will run a coding agent on it, give the model a longer memory for text. Ollama's
   [context length page](https://docs.ollama.com/context-length) asks for at least 64000 tokens for agents and coding
   tools:

   ```sh
   OLLAMA_CONTEXT_LENGTH=64000 ollama serve
   ```

   If Ollama already runs in the background, set the variable as
   [Ollama's server settings](https://docs.ollama.com/faq#how-do-i-configure-ollama-server) describe for your system.

Any other server that speaks the OpenAI interface works too. WireCat needs its address, and each server's docs give
the default one:

| Server | Address to enter in WireCat |
| --- | --- |
| [Ollama](https://docs.ollama.com/api/openai-compatibility) | `http://127.0.0.1:11434/v1` |
| [LM Studio](https://lmstudio.ai/docs/developer/openai-compat), server started in the Developer tab | `http://localhost:1234/v1` |
| [llama.cpp server](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md), `llama-server` | `http://127.0.0.1:9931/v1` |
| [vLLM](https://docs.vllm.ai/en/latest/getting_started/quickstart/), `vllm serve` | `http://localhost:8000/v1` |

Use the port your server really listens on, and keep the address on this computer, such as `127.0.0.1` or
`localhost`.

## 2. What is already local [#2-what-is-already-local]

You do not need to change anything for these:

- **Search by meaning** and **picture search** run on models that WireCat downloads to your computer.
- **Meeting recordings** are transcribed on your computer.
- **Chat linking, reply drafts and reading text in attachments with an AI model** stay off until you set a model for
  them yourself.

## 3. The AI model in Settings [#3-the-ai-model-in-settings]

This model reads the text in pictures you add to memory, and can read sources for facts. In WireCat Desktop open
**Settings**, **AI model**. See [AI model and better search](/llms.mdx/docs/desktop/ai-model/content.md) for every field.

**With Ollama:**

1. As **Provider**, choose **Ollama on this computer (beta)**, and pick the model from the list.
2. Tick **Read pictures with this model** if you want it to read pictures. The note under it says the pictures stay on
   this computer.
3. Save and click **Check**.

The same from a terminal. Here `--allow-remote` is what lets pictures go to the model, even though this model is on
your computer:

```text
memo models ai set ollama --model llava --allow-remote
```

**With another server on your computer:**

1. As **Provider**, choose **OpenAI or compatible**.
2. In **Address (base URL)**, enter your server's address from the table above.
3. Type the model's name as your server shows it.
4. In **API key**, type any value. WireCat asks for a key even for a server on your computer, and the server ignores
   it.
5. Save and click **Check**.

## 4. Reading for facts [#4-reading-for-facts]

In **Settings**, **Reading for facts**, choose **AI model**, which is now your local model, or **Ollama**. Each reader
says whether the text leaves this computer. Claude Code and Codex send it to their own model. Click **Use this
reader**. [Add to memory](/llms.mdx/docs/desktop/memory/content.md#add-to-memory) then suggests facts with that model.

## 5. Search harder and rewriting records for search [#5-search-harder-and-rewriting-records-for-search]

Both helpers are off by default. They are in Memo's details: on the Tools page, click **Manage** on Memo's row.

**Search harder** reorders the top results of one search when you press the button. Click **Allow**, choose
**OpenAI**, type your local model's name in **Model**, and enter `http://127.0.0.1:11434/v1` in **Address of another
OpenAI-compatible service**. Then agree with the button that names the provider. It waits only 5 seconds for an answer, so a model that runs without a graphics card will likely
run out of time, and the search keeps its own order.

**Rewrite records for better search** adds other words to your notes, documents and emails. WireCat Desktop offers only
cloud models for it, so leave it off there. From a terminal you can point it at your local server:

```sh
memo models consents set openai --model <model> --base-url http://127.0.0.1:11434/v1 --yes
```

Search harder takes the same address from a terminal:

```text
memo models search-consents set openai --model <model> --base-url http://127.0.0.1:11434/v1 --yes
```

## 6. Voice messages [#6-voice-messages]

By default the messenger makes the text from its own copy of a voice message, when it has a speech service, such as
Telegram Premium. To make the text on your computer with a model you downloaded, add `--local`:

```sh
tg <profile> messages transcribe <chat> <message> --local
```

When your agent asks for a transcription, it may not add `--local`. Ask it to, or use the command yourself.

## 7. Your coding agent on a local model [#7-your-coding-agent-on-a-local-model]

The agent you connect to WireCat over MCP or in a terminal reads your messages and notes through WireCat, then sends
them to its own model. To keep that local, run the agent on your local model. WireCat Desktop works with Claude Code,
Codex, Cursor, Gemini CLI, OpenCode, Hermes and OpenClaw. [Connect your AI agent](/llms.mdx/docs/agents/content.md) shows how to connect
one.

Ollama starts five of them on a local model with one command, `ollama launch <agent>`, where the agent is `codex`, `claude`, `opencode`,
`hermes` or `openclaw`. Ollama's pages for
[Codex](https://docs.ollama.com/integrations/codex), [Claude Code](https://docs.ollama.com/integrations/claude-code),
[OpenCode](https://docs.ollama.com/integrations/opencode), [Hermes](https://docs.ollama.com/integrations/hermes) and
[OpenClaw](https://docs.ollama.com/integrations/openclaw) describe it. When it asks for a model, pick one you
downloaded, not one that ends in `cloud`.

```sh
ollama launch codex
```

Without `ollama launch`:

- **Codex.** [Codex's advanced configuration](https://learn.chatgpt.com/docs/config-file/config-advanced) describes `--oss`
  for a local Ollama or LM Studio, and `oss_provider` in `~/.codex/config.toml` to choose one by default. Ollama's page
  gives `codex --oss -m <model>`.
- **Claude Code.** Ollama's [Claude Code page](https://docs.ollama.com/integrations/claude-code) sets
  `ANTHROPIC_BASE_URL=http://localhost:11434`, `ANTHROPIC_AUTH_TOKEN=ollama` and an empty `ANTHROPIC_API_KEY`, then
  runs `claude --model <model>`. Claude Code's settings can override these variables, so check with `/status`.
  [LM Studio](https://lmstudio.ai/docs/developer/anthropic-compat) describes the same with its own address. Anthropic
  [does not support](https://code.claude.com/docs/en/llm-gateway) Claude Code on models other than Claude.
- **OpenCode.** [OpenCode's providers page](https://opencode.ai/docs/providers/) says it finds Ollama, LM Studio and
  vLLM on this computer by itself. Start the server, then choose the model in OpenCode.
- **Hermes.** Run `hermes model` and choose a custom endpoint with your server's address, as
  [Hermes' providers page](https://hermes-agent.nousresearch.com/docs/integrations/providers) shows.
- **OpenClaw.** Run `openclaw onboard` and choose Ollama, then **Local only**, as
  [OpenClaw's Ollama page](https://docs.openclaw.ai/providers/ollama) shows. Use Ollama's address without `/v1`.

**Cursor** and **Gemini CLI** have no setup in their docs that runs the agent itself on a model on your computer.

Every one of these docs asks for a model that can call tools, and a context of at least 64000 tokens. A small model
on a laptop answers slowly and makes more mistakes with WireCat's commands than a cloud model.

## 8. Digests and the mail agent [#8-digests-and-the-mail-agent]

[Digests](/llms.mdx/docs/desktop/digests/content.md) and the mail agent in WireCat Desktop run Claude Code on Anthropic's model. They start
without your Claude Code settings, so the steps above do not reach them. Today they send the mail and messages they
read to Anthropic. If you want everything local, do not turn them on.

## Checklist: everything local [#checklist-everything-local]

1. Ollama or another model server runs on this computer, and Ollama's cloud models are off.
2. **Settings**, **AI model** is Ollama or an address on this computer.
3. **Reading for facts** uses the AI model or Ollama.
4. **Search harder** is off, or points at your local server. **Rewrite records for better search** is off, or set from
   a terminal to your local server.
5. Voice messages are transcribed with `--local`.
6. Your agent runs on a local model, or you connect no agent.
7. Digests and the mail agent are off.

## What still leaves your computer today [#what-still-leaves-your-computer-today]

- Digests and the mail agent, which run Claude Code on Anthropic's model.
- Anything your agent reads, when the agent runs on a cloud model.

WireCat is working on one switch that keeps every feature on your computer and shows what is local and what is not.

## Next [#next]

[AI model and better search](/llms.mdx/docs/desktop/ai-model/content.md) explains each setting in more detail.