# 1.26315 — 11 November 2026 You can now add a chat model that isn't in Gezel's catalog, from a Hugging Face link or a `.gguf` file already on your computer. Gezel checks it before anything downloads and tries it once it is here. Apps in your projects now update themselves when an update changes nothing of yours. The title bar gains Back and Forward, and a few names change: the manual is **Help**, and craftbooks are **Plans**. If you read only one thing: open your engine's page in **Settings** and look at **Add a model…**. ## Add a model that isn't in the catalog The catalog is a short list of models Gezel has tested. Until now, it was the only list you could choose from. On your engine's page in Settings (**This PC (llama)**, or **This Mac (Apple MLX)** on a Mac), choose **Add a model…**, then **From Hugging Face or a file…**. Paste a link to a model on [Hugging Face](https://huggingface.co), or choose a `.gguf` file you already downloaded with another app. Gezel runs GGUF models on every computer, and MLX models on Apple silicon Macs. Nothing downloads until you have read a review of the model and chosen **Download**. The review checks that anyone can download it and every file has a checksum, that it is a chat model and not an add-on part, that this version of Gezel's engine can run it, that it brings its own conversation format, and that it fits your computer's memory and disk. Gezel downloads only the model's data files, never program code, and turns away an MLX model whose settings ask to run code. If a check fails, the review says why, and a smaller size of the same model often fits where the largest does not. Once the model is here, Gezel tries it: it starts the model, asks for a short reply, checks that the reply reads as language, and asks it to use a tool. The result appears beside the model in the engine's **Models** list. Gezel doesn't pick a model that failed for work on its own, but you can still choose it yourself. A model you add uses the settings of a tested catalog model built from the same base model when there is one, and the settings its author recommends when there isn't. [Local models and tiers](https://gezel.com/docs/local-models-and-tiers/) has more on choosing one. ## Apps update themselves A project made from an app now moves to the app's newer versions on its own, as long as the update changes nothing of yours. Your data, your About and mission text, plans you edited and settings you changed stay as they are, and an update never runs in the middle of work. Some updates wait for you instead. One that would replace a script you changed, add a crew member, start something on a schedule, or write into a folder Gezel has no permission for shows **An update is ready** in the project's **Settings**. The dialog lists what changes and what stays the same, and a script it replaces is saved first in the project's artifacts, under `app-updates`. An app you added from a file, and a project you set to stay on its version, always wait for you. ## Back and Forward, and new names The title bar has **Back** and **Forward** keys, like a browser, and the arrow beside them lists the views you visited. Alt+Left and Alt+Right work too (Cmd+[ and Cmd+] on a Mac), and so do the side buttons on a mouse. Some names change: - The built-in manual is **Help**, everywhere it appears. - Craftbooks are **Plans** in the navigation. - In a project where one crew member does all the work, they are its **Main crew member**. A crew's lead is still the **Crew Lead**. - The places you use to build with Gezel fold into a group called **More** at the bottom of the navigation. It starts open if you told Gezel you write software or turned on advanced features. The title bar and a project's status bar are quieter for everyone else, and **This PC** stays in the title bar while a local model is in use. The navigation now sits on the left unless you chose a side. If you liked it on the right, move it back in **Settings → General → Sidebar position**. While Gezel starts, the window shows its real frame and navigation from the first moment, so nothing jumps when the app takes over. ## One row per request in Activity **Activity** shows each thing you asked for as one row, with the work it caused folded inside. A finished result can be put away: it moves under **Finished earlier** instead of leaving the list. **Settings** is also shorter, with its less-used pages gathered under **More settings**. ## Smaller things - You can type and send while a model is still loading, including during the first download. The message waits, and the composer shows **Sends when ready**. - In a reply, a sentence the fact check could not find in any source is highlighted, instead of being mentioned only in the line under the reply. - **Save a copy…** suggests the work's title as the file name, instead of `report.docx`. - **Settings → This PC (llama)** and the Mac's MLX page have a **Parallel workers** setting: how many chats one loaded model answers at once. Automatic runs as many as fit in memory. - On Claude Sonnet 5.5 and GPT-6.1 Sol, a plan now runs as one piece of work toward its goal. One crew member does it from start to finish, with the plan's steps as advice, and the plan's checks still decide when it is done. - When you don't choose, a new project starts with a crew or with one main crew member working alone, depending on the model you use. - When a cloud service refuses requests for coming too fast, Gezel slows down and keeps going, instead of letting each request time out. - A picture in chat has **Make a video** in its menu when a video model that animates pictures is installed. When the model can't use a picture, Gezel says so instead of making a clip from words alone. - With ffmpeg installed, a finished video also arrives as an MP4 with a poster picture. - The video model list marks models too large for this computer. - An on-device engine that gets stuck repeating one character is now noticed and restarted. Before, it could fill every reply with "/" for hours. - The first search on a new install no longer waits for the search model to download. - When work passes to a crew member whose model is still making room in memory, the handoff waits, instead of pausing the task. - Choosing an MLX model on a computer without Apple silicon now says why it can't run. - Every CSV file with rows in it, even a small one, now becomes a table Gezel can query and add up correctly. A folder of code is the exception, where a small CSV is usually test data. - The morning review of the night's work handles projects with very many tasks. - **Turn it on in Project Settings** opens the right page, and creating a Meester, Scribe, Librarian or Inspector in Settings now shows what went wrong when it fails. - On Linux, the `.deb` and `.rpm` give Gezel's shared engine access to the graphics card themselves, so local models can use an AMD card right after you install. ## Privacy and security - **Camera pictures and paired computers.** A paired server's video engine now follows the same rule as its picture engine: unless you let cloud AI see a camera's pictures, it does not get them. More in [Cameras and lookouts](https://gezel.com/docs/cameras-and-lookouts/). - **App updates need you.** A gezel can't approve an app update or apply one that waits for you. ## In the terminal - `gezel integrations address` shows where other programs reach this Gezel's models. `gezel integrations ollama on` or `off` opens or closes the port without a key that apps looking for Ollama use on this computer. `gezel integrations list`, `approve` and `deny` handle apps asking to connect. - `gezel project edits on` or `off` lets Gezel change files in the current project's folder, or stops it. - `gezel do --allow-edits` turns edits on for the project's folder before it starts the plan, and they stay on for that project. - `gezel craftbook show ` shows what a plan does, what it takes, and the command that starts it. - `gezel task wait ` waits for a task and says where its result is. `--out ` copies what it made into that folder. ## New in Help - A new [scorecard round](../technical/scorecard-rounds/2026-11-03-windows-saluki-native-0-1-57.md) measures a 27B model on a laptop with a 12 GB graphics card, with part of the model in system memory. ## Things to see and do after update - Open **Settings → This PC (llama)** (or the MLX page on a Mac) and try **Add a model…** with a model you have wanted to run. - If you prefer the navigation on the right, set **Settings → General → Sidebar position**. - If you use a project made from an app, look in its **Settings** for **An update is ready**.