Custom API enabled on Lio AI and Group Chat

[snakedev]

The Custom (OpenAI-compatible) is not enabled on Lio AI and the Group chat..

Tasks don't seem to get done.

And Autopilot doesn't seem to work correctly either

I've created multiple agents with a team structure... the CEO says he assigned tasks to the other agents but nothing seems to work.. no tasks assigned, other agents are not being spawned...

Am i using it wrong trying to automate?

[talentsache]

Thanks for the report. Let me address each point:

Custom API on Lio AI / Group Chat:
Which custom endpoint are you using? (LM Studio, Ollama, text-generation-webui, or something else?) Some endpoints don't support function calling / tool use which Lio AI and Group Chat rely on. The model needs to handle the ReAct loop for tool execution.

Best results with custom endpoints: use models that support tool calling natively (Qwen 2.5, Mistral, Llama 3.1+). Smaller models (7B and below)
often struggle with the ReAct pattern and fail silently.

Autopilot / Multi-Agent:
The Swarm system (multi-agent with CEO/team structure) is currently in Beta and in the time you were testing in alpha. Known limitation: the CEO agent can plan and delegate, but spawning sub-agents requires all agents to use the same provider. If your custom endpoint has rate limits or slow response times, the delegation chain breaks.

What might help:

1. Start with a single agent first. Get one task working end-to-end.
2. Use a faster provider for Autopilot (Groq or OpenRouter free tier work well for testing).
3. Check Settings > AI Providers: is the custom endpoint set as default for ALL features or only for Chat?

Quick test: Open a new Chat (not Lio AI, not Group Chat) and ask something simple like "What time is it?" or "Create a file called test.txt". Does the agent execute it? If yes, the endpoint works for Chat but might not support the tool schema that Lio AI / Group Chat need.

What version are you on? v8.0.0 just shipped with improvements to tool execution and timeout handling.

Let me know your setup (provider, model, OS) and I can help debug.

And sorry for the late reply, seems like I didn't get any notification about your thread.

[talentsache]

Adding to my previous reply:

Actually I think I see what you're doing - you created multiple agents (like a CEO, developer, designer etc.) and expected them to work together on tasks automatically. That's a great use case but let me clarify how the system works right now:

Chat Agents: When you start a chat with an agent (like your CEO), that agent can delegate tasks to other agents in the same conversation. So if you tell the CEO "research X, write Y, and design Z", the CEO should split the work and delegate. But this only works within Chat, not across separate sessions. (Multitask-Job, try it writing inside chat something like 'start a multi-task job with the goal...' give directly a bigger tasks through the chat)

Autopilot: This is a different system. Autopilot uses your DEFAULT provider and DEFAULT persona (set in Settings > General). It doesn't use the custom agents you created. It runs scheduled tasks independently.

So right now there's a gap: your custom agents work in Chat but Autopilot doesn't know about them.

You seem like a power user who's pushing exactly the right boundaries. I'm going to prioritize adding custom agent support to Autopilot in the next update so you can assign specific agents to specific automated tasks.

For now, the best workflow: start a Chat with your CEO agent, give it a complex multi-step task, and let it delegate within that conversation. For recurring tasks, use Autopilot with a detailed persona description in Settings > that covers what you need.

Thank you for your feedback!

[v33-kind]

@talentsache

side note : agent swarm could possibly solve this with some tweaking.
have a few ideas on this, will test it first.

[talentsache]

@v33-kind Good thinking - Swarm is exactly where multi-agent orchestration belongs long-term. Would love to see what you come up with. If you test anything, share your findings here or open a new discussion.

[snakedev]

I use multiple instances of codex-lb locally which is openai compatible.

each codex-lb is connected tu multiple openai oauth accounts and is going through them in round robin way.

i use multiple codex-lb installs so it doesn’t slow down everyone.. each tool i use is connected to a different one through api.

with agent swarm, i would like it to have it go through the different custom integrations..

Talk to the CEO tell it what to do, and use the other agents to do the tasks.. instead of spawning new ones with minimum system prompt and knowledge base, to use the ones i create separately with much better system prompts.

same for Lio AI, to be able to use my custom api integrations (codex-lb)

i have a lot of accounts so i can push it to the maximum to see what it can do..

last 7 days alone, i got through over 2 billion tokens on 3 different codex-lb installs.

[snakedev]

Plus 17 notion ai agents… everyone working on https://healthsecrets.com

[snakedev]

Also, one issue i found is that if i have something else running on port 3000, skales still starts normally but it’s showing whats already on port 3000 without attempting to use another port or showing an error

[snakedev]

Something else i would love to see, is browser use and computer use.. was doing some tests with rtrvr.ai today..

use case for me, i have a notion database, a marketing calendar built and updated by 17 notion ai agents.

it has multiple columns, Platform (medium, quora, substack etc.. about 14 of them), Status (Ready to Publish), Draft Link (link to notion subpage with the content that goes on the platform) - agents looks through the database, finds the next row with Medium and Ready to Publish, opens the draft link page - grabs the title, content, tags, image prompts, opens medium and creates the post - i have some of the working like this with a local automation that grabs the image prompts, opens google flow, generates the image, download it, rename it to the suggested name, moves it to an article folder with the article id, which it creates automatically - if all images are created, goes to medium, creates a new draft, adds the title, content in inserts the images in the suggested placement of each, clicks continue, adds the post tags, save it as a draft ( i can make it publish it directly..

[talentsache]

Hey Stefan,

Thanks for the detailed breakdown of your setup! I appreciate your 'Ultimate-Stresstest' on Skales!

Port 3000 conflict: confirmed bug, goes into the next patch. Skales should auto-detect and fail gracefully or pick the next free port. Right now it silently shows whatever is already on 3000. Thanks for catching this.

Custom API on Agent Swarm: your use case makes total sense - having each agent connected to a different codex-lb instance with its own system prompt and knowledge base, instead of spawning new ones from scratch. That's being considered for the next update.

Using pre-configured agents instead of spawning minimal ones: agreed. You put work into building agents with proper system
prompts, they should be reusable across Swarm and other features.

Lio AI with custom API integrations: currently uses the default provider. Supporting custom endpoints for Lio is on the list based on your feedback.

2 billion tokens in a week across 3 codex-lb instances plus 17 Notion AI agents on healthsecrets.com 🤯 that's serious infrastructure. This kind of real-world stress testing is exactly what I need to hear about.

Your input has directly shaped the next update - you'll get the contribution credit you deserve!

Keep pushing it and let me know where it breaks.

Updates coming soon.

[snakedev]

Thank you for listening and considering all this.

let me know once you implement all this and the lio ai with the custom provider and i will stress test the hell out of them and let you know what other bugs i find and what other features would be usefull.

[snakedev]

I can give you a tip how i use so many tokens for cheap but somewhere is private

[snakedev]

I sent you an email

[talentsache]

@snakedev Good news - Lio AI custom provider and Autopilot agent assignment are both going into v8.0.1 which is being built right now. I'll ping you here once it drops.

Thanks for the email, checking it now.

[snakedev]

@talentsache i'm on v 8.0.0 and see its available for Lio AI

I have some ideas for Lio AI:

• instead of just 3 agents (Architect, Reviewer, Builder) - to allow a custom configuration use of the custom agents for advanced users... it-s not smart enough to do something good on its own yet -

And many times, not even deploying the html site in the preview:

I have a lot created with advanced prompts and they know how to talk to each other, if allowed to spawn each other....

• to create, build and run the project locally automatically... just like i would do in vscode, cursor or any other code builder (same thing for Autopilot)

• let it run in the background so i can go back to chatting with agents - if i switch tabs, it stops..

• multiple Custom providers and to allow them in Vision Provider as well

More advanced customization for Lio AI - on top of allowing to select multiple agents with custom prompts, to allow the use of Skills and custom skills as well.

• github integration... if i want to work on a github project - to have it analyze it and make improvements to it, adding new features etc... then push and commit

Also, few bugs i just encountered:

• if it finishes a building something, you can't give me instructions to make more changes... after a couple of tries, the input box doesn't work anymore - you have to close and restart the windows app...
• Autopilot - executed a task, said it created a file, but it didn't.. nothing was saved..

I was thinking of forking it and attempting to implement a lot of these new features, but it might not be a clean code, althought, i can add a skill in vscode to make sure gpt-5.4 outputs clean code and well commented.

The way i developed healthsecrets.com - i spent 2-3 weeks, with 2-3 billion tokens per week - at night, i was telling Codex or vscode to create a large list of improvements and features, create 50-60 tasks for itself and go through them all night... sometimes i was having 2 of them (Codex and VScode) - each connected to a different codex-lb instance with different openai Oauth accounts, work in parallel all night - one working at the front end, another working at the backend... and i built a custom CMS with pretty much everything it needs and more and fully optimized - still improving it

[talentsache]

@snakedev thanks as always for the detailed feedback and screenshots.

The bugs you found (input box freezing after build, Autopilot not saving files) are noted and being worked on.

The feature ideas are good input. Some are already on the roadmap, others need more thought.

v8.0.1 just shipped with background execution support and several fixes. v8.0.2 already in the pipeline.

Please update and let me know how it goes.

[snakedev]

Thanks. I’ve forked it and added a bunch of features including multiple custom providers, ability for leo ai to select any built agent for every step, not just the default ones, mcp support, ability for agents to use other agents, not just spawn new ones and a lot more that i need to test this week, but i have my 3 year old sone at me and he doesn’t let me near the pc 😅 Once i test everything i will push everything and you can access my forked version to test it Sincerely, [image: photo] Stefan Vintila Tech Guru at Kanium E ***@***.*** ***@***.***> W https://kanium.com <https://kanium.com> Skype stefan.bartender <http://facebook.com/stefan%5C.vintila> <http://us.linkedin.com/in/stefanvintila> <http://twitter.com/SnakeMarketing> <http://hatscripts.com/addskype?stefan.bartender> <http://wiseintro.co/stefanv> [image: App Social Buttons Image] <https://www.facebook.com/stefan.vintila> [image: App Social Buttons Image] <https://www.linkedin.com/in/stefanvintila/> "Life would be tragic if it weren't funny." - Stephen Hawking.

…

On Sun, 29 Mar 2026 at 19:42, Mario Simic ***@***.***> wrote: @snakedev <https://github.com/snakedev> thanks as always for the detailed feedback and screenshots. The bugs you found (input box freezing after build, Autopilot not saving files) are noted and being worked on. The feature ideas are good input. Some are already on the roadmap, others need more thought. v8.0.1 just shipped with background execution support and several fixes. v8.0.2 already in the pipeline. Please update and let me know how it goes. — Reply to this email directly, view it on GitHub <#34?email_source=notifications&email_token=ABAIBZIUHC6AVARWMZXXZQT4TFOBFA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNRTGY4DSNJYUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-16368958>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABAIBZPOS5DRWI2RMJ2OLC34TFOBFAVCNFSM6AAAAACW273VU6VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMMZWHA4TKOA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[snakedev]

Also a way to always find a free port when 3000 is taken Sincerely, [image: photo] Stefan Vintila Tech Guru at Kanium E ***@***.*** ***@***.***> W https://kanium.com <https://kanium.com> Skype stefan.bartender <http://facebook.com/stefan%5C.vintila> <http://us.linkedin.com/in/stefanvintila> <http://twitter.com/SnakeMarketing> <http://hatscripts.com/addskype?stefan.bartender> <http://wiseintro.co/stefanv> [image: App Social Buttons Image] <https://www.facebook.com/stefan.vintila> [image: App Social Buttons Image] <https://www.linkedin.com/in/stefanvintila/> "Life would be tragic if it weren't funny." - Stephen Hawking. On Sun, 29 Mar 2026 at 22:13, Stefan Vintila ***@***.***> wrote:

…

Thanks. I’ve forked it and added a bunch of features including multiple custom providers, ability for leo ai to select any built agent for every step, not just the default ones, mcp support, ability for agents to use other agents, not just spawn new ones and a lot more that i need to test this week, but i have my 3 year old sone at me and he doesn’t let me near the pc 😅 Once i test everything i will push everything and you can access my forked version to test it Sincerely, [image: photo] Stefan Vintila Tech Guru at Kanium E ***@***.*** ***@***.***> W https://kanium.com <https://kanium.com> Skype stefan.bartender <http://facebook.com/stefan%5C.vintila> <http://us.linkedin.com/in/stefanvintila> <http://twitter.com/SnakeMarketing> <http://hatscripts.com/addskype?stefan.bartender> <http://wiseintro.co/stefanv> [image: App Social Buttons Image] <https://www.facebook.com/stefan.vintila> [image: App Social Buttons Image] <https://www.linkedin.com/in/stefanvintila/> "Life would be tragic if it weren't funny." - Stephen Hawking. On Sun, 29 Mar 2026 at 19:42, Mario Simic ***@***.***> wrote: > @snakedev <https://github.com/snakedev> thanks as always for the > detailed feedback and screenshots. > > The bugs you found (input box freezing after build, Autopilot not saving > files) are noted and being worked on. > > The feature ideas are good input. Some are already on the roadmap, others > need more thought. > > v8.0.1 just shipped with background execution support and several fixes. > v8.0.2 already in the pipeline. > > Please update and let me know how it goes. > > — > Reply to this email directly, view it on GitHub > <#34?email_source=notifications&email_token=ABAIBZIUHC6AVARWMZXXZQT4TFOBFA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNRTGY4DSNJYUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-16368958>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/ABAIBZPOS5DRWI2RMJ2OLC34TFOBFAVCNFSM6AAAAACW273VU6VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMMZWHA4TKOA> > . > You are receiving this because you were mentioned.Message ID: > ***@***.***> >

[snakedev]

And any agent can be used on the Autopilot feature as well Sincerely, [image: photo] Stefan Vintila Tech Guru at Kanium E ***@***.*** ***@***.***> W https://kanium.com <https://kanium.com> Skype stefan.bartender <http://facebook.com/stefan%5C.vintila> <http://us.linkedin.com/in/stefanvintila> <http://twitter.com/SnakeMarketing> <http://hatscripts.com/addskype?stefan.bartender> <http://wiseintro.co/stefanv> [image: App Social Buttons Image] <https://www.facebook.com/stefan.vintila> [image: App Social Buttons Image] <https://www.linkedin.com/in/stefanvintila/> "Life would be tragic if it weren't funny." - Stephen Hawking. On Sun, 29 Mar 2026 at 22:14, Stefan Vintila ***@***.***> wrote:

…

Also a way to always find a free port when 3000 is taken Sincerely, [image: photo] Stefan Vintila Tech Guru at Kanium E ***@***.*** ***@***.***> W https://kanium.com <https://kanium.com> Skype stefan.bartender <http://facebook.com/stefan%5C.vintila> <http://us.linkedin.com/in/stefanvintila> <http://twitter.com/SnakeMarketing> <http://hatscripts.com/addskype?stefan.bartender> <http://wiseintro.co/stefanv> [image: App Social Buttons Image] <https://www.facebook.com/stefan.vintila> [image: App Social Buttons Image] <https://www.linkedin.com/in/stefanvintila/> "Life would be tragic if it weren't funny." - Stephen Hawking. On Sun, 29 Mar 2026 at 22:13, Stefan Vintila ***@***.***> wrote: > Thanks. > > I’ve forked it and added a bunch of features including multiple custom > providers, ability for leo ai to select any built agent for every step, not > just the default ones, mcp support, ability for agents to use other agents, > not just spawn new ones and a lot more that i need to test this week, but i > have my 3 year old sone at me and he doesn’t let me near the pc 😅 > > Once i test everything i will push everything and you can access my > forked version to test it > > Sincerely, > > [image: photo] > Stefan Vintila > Tech Guru at Kanium > E ***@***.*** ***@***.***> > W https://kanium.com <https://kanium.com> > Skype stefan.bartender > <http://facebook.com/stefan%5C.vintila> > <http://us.linkedin.com/in/stefanvintila> > <http://twitter.com/SnakeMarketing> > <http://hatscripts.com/addskype?stefan.bartender> > <http://wiseintro.co/stefanv> > [image: App Social Buttons Image] > <https://www.facebook.com/stefan.vintila> [image: App Social Buttons > Image] <https://www.linkedin.com/in/stefanvintila/> > > "Life would be tragic if it weren't funny." - Stephen Hawking. > > > On Sun, 29 Mar 2026 at 19:42, Mario Simic ***@***.***> > wrote: > >> @snakedev <https://github.com/snakedev> thanks as always for the >> detailed feedback and screenshots. >> >> The bugs you found (input box freezing after build, Autopilot not saving >> files) are noted and being worked on. >> >> The feature ideas are good input. Some are already on the roadmap, >> others need more thought. >> >> v8.0.1 just shipped with background execution support and several fixes. >> v8.0.2 already in the pipeline. >> >> Please update and let me know how it goes. >> >> — >> Reply to this email directly, view it on GitHub >> <#34?email_source=notifications&email_token=ABAIBZIUHC6AVARWMZXXZQT4TFOBFA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNRTGY4DSNJYUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-16368958>, >> or unsubscribe >> <https://github.com/notifications/unsubscribe-auth/ABAIBZPOS5DRWI2RMJ2OLC34TFOBFAVCNFSM6AAAAACW273VU6VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMMZWHA4TKOA> >> . >> You are receiving this because you were mentioned.Message ID: >> ***@***.***> >> >

[talentsache]

Stefan, impressive speed with the 3 year old on your lap.

Since my main branch is several versions ahead of what's on GitHub, a direct merge won't work. But I'd love to review your implementations and use them as reference for the main branch.
Especially interested in how you handled MCP and agent-to-agent.

Send me the fork when ready, I'll go through your code and integrate what works into the next release with full credit.

Mario

[snakedev]

I went a bit crazy with the features in my forked version...

[talentsache]

@snakedev This is impressive. You basically built the enterprise multi-agent layer on your own fork. Browser Playbooks and Org Chart with delegation modes - that's exactly where Skales is heading.

v8.0.2 just shipped with a ton of stability fixes plus
Discover Feed, new themes, and more. Might be worth syncing your fork.

Let me know if you want to chat about aligning some of these features with mainline.

[v33-kind]

@snakedev this is really impressive, and i love it, forking it. also followed you.
@talentsache - i have a few ideas on this!

i live for this kind of thinking guys!!!!!

[snakedev]

Still adding a lot more

Lio Ai and Autopilot can now use Companies & Teams foe every step, instead of just 1 agent for each... When you give a task to autopilot, Lio or the CEO agents, it starts a team meeting with the departments heads to plan everything, then they start to deploy to specialized agents. They can also use Computer-use directly and browser playbooks ( Lio ai, autopilot, teams or agents ) - so a marketing team can deploy marketing agents that will use the browser (external profile - Chrome or internal ) to access different sites and create posts, like medium, substack, linkedin and pretty much anything...

Google stich integration so the teams can generate designs and have engineering implement them.. and google flow through browser user so lio, autopilot and pretty much everything else will ask for design directions, prompts, and use the browser to generate images and designs, then hand them to engineering team to implement.

Multiple Openrouter accounts, multiple grok accounts, notion integration, google drive, sheets and docs integration... and github... working on having lio ai, autopilot and the teams be able to deploy to github directly.

Company memory, team memory, agent memory and auto evolution so they learn to do better... borrowed the best features from multiple sources, like hermes agent, superpowers, paperclipai and more...

And the whole company is used in lio ai now.. still working some kinks to recognize the department heads but this went from using 1 agent for each step to using departments or the whole company for each step (Architechture, Review, Build)

[snakedev]

I think i've built a monster here...

It now properly delegates to departments, generates designs with google stitch, images with google flow.

[talentsache]

@snakedev Thanks for sharing your work. Your fork screenshots gave the push to ship these features earlier than planned. Multi-agent teams, organization charts, and team delegation are now native in v9.0.0 alongside Codework, Computer Use, Calendar Sync, Agent Skills import, 7 integrations, and more.

Closing this thread to keep discussions focused on the mainline release. If you want to discuss v9.0.0 features or report bugs, the Issues tab is open.

🦎