A lot has shipped since the 2.1.7 post. Agents now remember what lasts about you between chats, check their own work while they run, and ask you less. Localoy updates itself from inside the app, and works in a browser tab as well. This post rounds up everything from 2.1.8 to 2.1.20.
New
- Memory. Agents keep what lasts about you between chats: your business, how you like to work, what you decided, and anything you ask them to remember. A new chat starts already knowing it. Say “like we talked about” and the agent finds the earlier chat. Every memory is on its own page in the agent’s settings, where you can see, edit and delete it, turn memory off, or share it with all your agents. Agents never keep a password, key, card or ID number.
- Updates from inside the app. Localoy fetches new versions in the background and checks they are really ours. When one is ready you see it at the top of the window; press Restart and it comes back on the new version. If an agent is working, the update waits.
- Localoy in your browser. Everything the app does also works in a tab at
http://127.0.0.1:2468. Install from Terminal withcurl -fsSL https://localoy.ai/install.sh | shon macOS or Linux, including on a server. - Quick decisions. A small, fast model answers the simple questions while your agents work: does this lead fit the plan, which button next in the browser. With it, agents leave out search results that are off-topic, check every page they read for text that tries to give them instructions, and check a helper’s answer against what it was asked. Each check acts only when it is sure.
- One yes for a whole site. When an agent asks to click or type on a site such as linkedin.com, one answer covers the rest of that chat. A schedule can keep it too.
- Agents remember websites, like where a button lives, and use it next time.
- Gmail. Agents can read, search and write drafts in your Gmail. They never send.
- Glow and crystal in 3D. In Vinci, any shape can glow and light what is around it, and crystal is clear glass that shows what is inside.
Changed
- One model list. Pick from Localoy’s models, each with its maker and a search box, or a model downloaded to your computer. A new chat keeps the model of the chat you opened it from.
- Credit instead of plans. No plans and no trial. Add Localoy credit when you want more; it never expires, and what it unlocks stays unlocked.
- Each agent sticks to its own skills. When a job needs another agent’s skill, it asks you first, for that chat only.
- A calmer chat. What an agent says between steps shows as two lines, long turns fold their earlier steps, and nothing jumps while it works.
- Chats you can find. An agent’s chats are grouped by Today, Yesterday and Earlier, with a find box once there are more than eight.
- Usage for today. The usage panel shows this chat, its agent and all agents for today, in money and tokens. The full history is on app.localoy.ai.
- Fewer “Keep going” stops. A turn that is still making progress keeps going, within your daily spend limit.
For teams
- API keys for every member. On app.localoy.ai, each member of an organization can make their own key and use Localoy models from their own scripts and tools at
https://api.localoy.ai/v1. A member can make one once an admin has set their monthly limit, and the key spends within it, only on the models the organization allows. Usage shows what came from the app and what came from each key.
Fixed
- Agents set up before 2.1.16 on a model Localoy no longer runs move to Localoy’s model by themselves, instead of every new chat being refused on its first message.
- A message sent while an agent works always gets a reply, and Stop now means pause: type “continue” to pick the work back up.
- A message never lands in an empty new chat when the chat list fails to load, and usage counts every call a turn made.
Already on 2.1.14 or later? Localoy updates itself. Otherwise, download 2.1.20 from the latest release.