I've been a software engineer/architect since 2000. I built Lucidos, a self-hosted personal agent platform. It has a workspace of files, an event-sourced memory, cron and event triggers, and it builds apps for you on demand and keeps them running. Here's a demo video it made of itself, role-playing a user: https://youtu.be/hNN5bL1nq7k
Backstory: Around a year ago I wanted to have my son create a game directly through a chat, and have it running as we iterated on it. The models were capable enough to do that, so why didn't anyone ship a harness to do it? Generally it bothered me how little AI agents could actually do for me. I started discussing with ChatGPT how to architect an agent that could do useful things with less friction. Then I had my third child and didn't think too much about it before February this year when I started using Claude Code a lot for work.
Claude Code was useful, but still just a coding agent, not creating the stuff I needed directly. But it was enough to start coding something more useful: Not long after Lucidos was born, and the first app it ran was a snake game that me and my son iterated on. We added Mario sound when eating the "apples", a high score list of "legends", replays and some funny music snippets (those needed a lot of iteration, but all AI prompted, from way back then with Opus 4.5). As a more mature platform Lucidos now has plugins and the snake game is available as a curiosity here:
https://github.com/lucidos-dev/plugins/tree/main/ouroboros (can be installed from Plugins in the platform)
Then I got Lucidos to run Claude Code in its own threads, in worktrees automatically (manually handling that is too much for me, I messed it up a lot). Development sped up, and got even better when I made it work as a mobile PWA as well. Then I could continue iterating on it in between life as a busy family man. Some (my wife) would (does) say I use my phone too much though.
The platform is event sourced, so everything you chat about is stored as events, from which it extracts memories. It has its own workspace of files for apps and artifacts. You can set up automations with cron triggers, or event triggers on any system or custom events. You can set up push notifications for them as well.
The platform sets you up for hyper specialized, personally customized apps, that compound as you give it more data sources. I find myself spinning up apps for everything these days, because it's so low cost. Some examples:
- Sauna app: Click a start button and it starts logging a sauna session, playing music on my Sonos speaker there for a configurable duration like 20 minutes. Why do I need this? The sessions get logged into my health data (oura, fitbit weight, blood samples etc) which is used by the Holistic doctor app.
- App handling finances/social services for my mom, as dad had to move to a retirement home. Tracking progress, finding phone numbers, drafting complaints, researching laws (there's a lot to do!). Continuously doing research using chat, and updating a structured app to track status, is a general pattern I've learned to appreciate. With Lucidos you can update your data in chat or custom interfaces. Always having both options is great - e.g. Shopping List app best when in the store, chat to research recipes and add to it.
- Sonos + Spotify app. Airplay through the phone is quite unreliable in my home (most homes?) so I run my own hyper specialized Music app that has the same integration to Sonos and Spotify that the agent has. From the app I can have Sonos say stuff like "What the fuck are you doing?" to my brother in law when he sits alone in our living room. That was fun.
- I also run the Lucidos CI as a trigger, orchestrating multiple coding agents to keep the code clean, and run a suite of e2e tests for all the clients, with a little dashboard app
What makes the Lucidos platform different?
- Everything lives in the same workspace, so you can discuss and edit all your data and apps in chat. This enables iteration at any point, you can do more research and change apps as you're chatting, based on new data or use cases. It's all available to the agent.
- Threads can spawn threads, as children or top level. This applies whether the thread is driven by the native Lucidos agent, Claude Code, or Codex. They can all spawn new threads.
- Credentials you configure are available in chat, automations and real time proxies serving the apps from the clients
- Anything created with Lucidos can be shared on a marketplace (github repo) and installed as plugins (not just apps, but Lucidos' automations and knowhow). So anyone can set up marketplaces with their apps (no need to be a developer) and they can be picked up by other Lucidos users with a click, available for them as well on desktop and mobile.
- There are convenience functions for productivity. Search Everywhere can open apps, files, settings etc. The agent can open anything as well (it can e.g. open your settings and change them while you watch - not by controlling the browser, but triggering Lucidos events. All events are relayed with SSE to all clients so they're always up to date). You can deeplink from anywhere to anywhere. Notifications also deeplink, so they can open your tracking app. The notification can be sent only if you forgot to log.
- It's free open source (MIT) and bring your own model
- It can edit itself while running. If anything bugs you, change it, also from mobile.
Since it's running locally it does store credentials unencrypted on your machine, so it's the same as using a .env basically. The agent can get your credentials if it needs to, but it's the engine that uses them. Your configured credentials never go through the chat/llm, unless you explicitly ask it to circumvent that and the model agrees (it's instructed to not do credentials in chat).
The best tested path is running from source using the web-dev.sh script, the DMG second. The other versions (install.sh, any linux) have been smoke tested by agents, not real users.
I'd love help testing:
- run it with a local setup of a frontier open source model, e.g. Kimi K3
- test the Android PWA (iOS is thoroughly tested, Android not)
- Linux install, end to end
What I don't have yet, that I miss and am looking into: voice (w the awesome new models for real conversation), and communication between nodes/multiplayer. I appreciate any feedback, what do you miss in the platform?
kennethtiller•2h ago
I've been a software engineer/architect since 2000. I built Lucidos, a self-hosted personal agent platform. It has a workspace of files, an event-sourced memory, cron and event triggers, and it builds apps for you on demand and keeps them running. Here's a demo video it made of itself, role-playing a user: https://youtu.be/hNN5bL1nq7k
Build from source (where you can also edit the platform itself): https://github.com/lucidos-dev/lucidos or install the DMG from https://lucidos.dev
Backstory: Around a year ago I wanted to have my son create a game directly through a chat, and have it running as we iterated on it. The models were capable enough to do that, so why didn't anyone ship a harness to do it? Generally it bothered me how little AI agents could actually do for me. I started discussing with ChatGPT how to architect an agent that could do useful things with less friction. Then I had my third child and didn't think too much about it before February this year when I started using Claude Code a lot for work.
Claude Code was useful, but still just a coding agent, not creating the stuff I needed directly. But it was enough to start coding something more useful: Not long after Lucidos was born, and the first app it ran was a snake game that me and my son iterated on. We added Mario sound when eating the "apples", a high score list of "legends", replays and some funny music snippets (those needed a lot of iteration, but all AI prompted, from way back then with Opus 4.5). As a more mature platform Lucidos now has plugins and the snake game is available as a curiosity here: https://github.com/lucidos-dev/plugins/tree/main/ouroboros (can be installed from Plugins in the platform)
Then I got Lucidos to run Claude Code in its own threads, in worktrees automatically (manually handling that is too much for me, I messed it up a lot). Development sped up, and got even better when I made it work as a mobile PWA as well. Then I could continue iterating on it in between life as a busy family man. Some (my wife) would (does) say I use my phone too much though.
The platform is event sourced, so everything you chat about is stored as events, from which it extracts memories. It has its own workspace of files for apps and artifacts. You can set up automations with cron triggers, or event triggers on any system or custom events. You can set up push notifications for them as well.
The platform sets you up for hyper specialized, personally customized apps, that compound as you give it more data sources. I find myself spinning up apps for everything these days, because it's so low cost. Some examples:
- Sauna app: Click a start button and it starts logging a sauna session, playing music on my Sonos speaker there for a configurable duration like 20 minutes. Why do I need this? The sessions get logged into my health data (oura, fitbit weight, blood samples etc) which is used by the Holistic doctor app.
- App handling finances/social services for my mom, as dad had to move to a retirement home. Tracking progress, finding phone numbers, drafting complaints, researching laws (there's a lot to do!). Continuously doing research using chat, and updating a structured app to track status, is a general pattern I've learned to appreciate. With Lucidos you can update your data in chat or custom interfaces. Always having both options is great - e.g. Shopping List app best when in the store, chat to research recipes and add to it.
- Sonos + Spotify app. Airplay through the phone is quite unreliable in my home (most homes?) so I run my own hyper specialized Music app that has the same integration to Sonos and Spotify that the agent has. From the app I can have Sonos say stuff like "What the fuck are you doing?" to my brother in law when he sits alone in our living room. That was fun.
- I also run the Lucidos CI as a trigger, orchestrating multiple coding agents to keep the code clean, and run a suite of e2e tests for all the clients, with a little dashboard app
What makes the Lucidos platform different?
- Everything lives in the same workspace, so you can discuss and edit all your data and apps in chat. This enables iteration at any point, you can do more research and change apps as you're chatting, based on new data or use cases. It's all available to the agent. - Threads can spawn threads, as children or top level. This applies whether the thread is driven by the native Lucidos agent, Claude Code, or Codex. They can all spawn new threads. - Credentials you configure are available in chat, automations and real time proxies serving the apps from the clients - Anything created with Lucidos can be shared on a marketplace (github repo) and installed as plugins (not just apps, but Lucidos' automations and knowhow). So anyone can set up marketplaces with their apps (no need to be a developer) and they can be picked up by other Lucidos users with a click, available for them as well on desktop and mobile. - There are convenience functions for productivity. Search Everywhere can open apps, files, settings etc. The agent can open anything as well (it can e.g. open your settings and change them while you watch - not by controlling the browser, but triggering Lucidos events. All events are relayed with SSE to all clients so they're always up to date). You can deeplink from anywhere to anywhere. Notifications also deeplink, so they can open your tracking app. The notification can be sent only if you forgot to log. - It's free open source (MIT) and bring your own model - It can edit itself while running. If anything bugs you, change it, also from mobile.
Since it's running locally it does store credentials unencrypted on your machine, so it's the same as using a .env basically. The agent can get your credentials if it needs to, but it's the engine that uses them. Your configured credentials never go through the chat/llm, unless you explicitly ask it to circumvent that and the model agrees (it's instructed to not do credentials in chat).
The best tested path is running from source using the web-dev.sh script, the DMG second. The other versions (install.sh, any linux) have been smoke tested by agents, not real users.
I'd love help testing: - run it with a local setup of a frontier open source model, e.g. Kimi K3 - test the Android PWA (iOS is thoroughly tested, Android not) - Linux install, end to end
What I don't have yet, that I miss and am looking into: voice (w the awesome new models for real conversation), and communication between nodes/multiplayer. I appreciate any feedback, what do you miss in the platform?