# Settings › AI models

> GLLmolfar can use your own language models for its written analyses. Add a connection, check it, and choose which model does which job. The section is marked **alpha**.

Describes GLLmolfar 0.9.22, checked 2026-11-01. Canonical page: https://www.getlooploop.com/gllmolfar/docs/settings-ai-models

- **Where:** Settings › AI models
- **Stage:** Alpha
- **Keys:** Encrypted on this computer, never shown again
- **Changes:** Saved at once, apart from the Settings form

## On the screen

![Settings › AI models in GLLmolfar with no connection yet: the intro and a button for each kind of provider](https://www.getlooploop.com/images/gllmolfar/docs/settings-ai-models.webp)

Numbered parts of the screenshot:

1. **Alpha and the intro.** **Connect your own models for analyses.** Where the page text goes, and how keys are kept.
2. **Add a connection.** One button per supported provider, each with its own mark, and **Other OpenAI-compatible** for any other server.
3. **Connections.** **No models connected yet. Add one above.** until the first one is added; then a card per connection.
4. **When it applies.** **Crawl and embedding settings apply to the next crawl.** The connections themselves are saved as you change them.
5. **Cancel and Save.** For the rest of the Settings form. Connections and routes do not wait for **Save**.

## Where the text goes

**A model on this computer or this network keeps the page text here; a cloud provider receives the text an analysis sends it. Keys are encrypted on this computer and never shown again.** Each card says which of the two it is: **On this computer or network**, or **Cloud:** and the host.

## A connection's card

Pressing a provider's button adds a card for it, with the provider's usual address filled in. Up to twelve connections can be kept.

| Part | What it shows or does |
| --- | --- |
| Header | The provider's mark and name, **On this computer or network** or **Cloud: host**, and **Remove**. Removing a connection also clears every job that used it. |
| Address (OpenAI-compatible, ends in /v1) | Where the server answers. It must start with `http://` or `https://`. |
| API key | Shown as **API key** where the provider needs one, **API key (if the server asks for one)** where it may not, and not shown where it is never used. A saved key reads **Saved — type to replace**. |
| State | **Not checked yet**, **Connected · models: N · N ms** with the date of the check, or the reason it failed. |
| Model list | After a successful check: the models the server offers. The model picked here is the one **Try a reply** asks. |
| Check | Saves the address and key as typed, then asks the server for its models. Enter in a field does the same. |
| Try a reply | After a successful check: asks the model for one short sentence and shows the reply with its time. A cloud card adds **A few tokens, billed to your key.** |
| What the model was thinking | When the model sent its reasoning with the reply, it folds under the reply. |

## Which model does what

Once a connection exists, a list of jobs appears under the cards. Each job takes one model from a checked connection, or **— not set —**. The note above it says: **Alpha: the choice is saved now; the analyses start using it in the next builds. Only checked connections are offered.**

| Job | What the app says it does |
| --- | --- |
| Topic names | Names each topic from its pages' titles and terms. |
| Explaining fixes | Why a problem matters and how to fix it, in plain words. |
| Page audit | A written summary of a single page's audit. |
| Reverse prompts | Judges a page's blocks and drafts their prompts. |
| Agent (test section) | Plans the work, ranks the tasks and checks its own findings. |
| Writing | Rewrites pages and writes articles in the agent's test section. |
| Fact checking | Checks every claim of a draft against its source, and judges pages flagged as risky. |
| Social posts | Posts, threads and articles for social networks. |

## States and messages

- **Connected** — the app says: "Connected · models: 12 · 140 ms" The server answered and listed its models. The connection can now be chosen for a job.
- **Needs a key** — the app says: "Reachable, but it needs your API key before it can answer." Type the key and press **Check** again.
- **Key refused** — the app says: "The key was refused." The server rejected the key.
- **Timeout** — the app says: "No answer within 10 seconds." The server did not answer the check in time.
- **Nothing there** — the app says: "Nothing is listening at this address. Is the server running?" Usually a local server that is not started.
- **No such host** — the app says: "This address does not resolve." The host name in the address is wrong.
- **Other** — the app says: "Could not connect: …" Any other failure, with its detail.
- **Only thinking** — the app says: "The model only thought and gave no answer within the limit. Thinking models need more room: it works in the agent, which allows long answers." **Try a reply** allows a short answer only.
- **Cut** — the app says: "The answer was cut by the length limit before any text came out." From **Try a reply**.
- **Empty** — the app says: "The server answered with empty text. Check that this model writes text (not an embedding model) and that it is loaded." From **Try a reply**.
- **No model** — the app says: "No checked model: connect one in Settings › AI models." Shown by an analysis that needs a model when none is checked.

## Related pages

- [AI](https://www.getlooploop.com/gllmolfar/docs/ai.md): The AI screen: an AI readiness score, which AI bots may get in and what they get, what blocks quoting, trust signals, llms.txt, Markdown for agents.
- [Agent · test](https://www.getlooploop.com/gllmolfar/docs/agent-test.md): The Agent · test screen: your own models plan tasks from the fix list, rewrite pages, write social posts and review risky pages; GLLmolfar measures the results.
- [Reverse prompts](https://www.getlooploop.com/gllmolfar/docs/reverse-prompts.md): Site › Reverse prompts: GLLmolfar reads a page into blocks, your AI agent describes what each does, and every block gets a reusable prompt and an example.
- [MCP tools for agents](https://www.getlooploop.com/gllmolfar/docs/mcp-tools.md): The MCP server inside GLLmolfar: how to connect an AI agent, and every tool it can call to crawl, read fixes, audit pages, search by meaning and save reports.
- [Settings › Embeddings](https://www.getlooploop.com/gllmolfar/docs/settings-embeddings.md): Settings › Embeddings: whether new crawls build embeddings, how much of each page the model reads, and whether it runs on the CPU or the GPU.