LobeHub - Your Chief Agent Operator for multi-agent work

Does CAO work with custom local models via something like Ollama, or only cloud APIs?

@arvinx This is only for the cloud version? Self-hosted, especially 2.2.0, is quite regressive.

- No functioning CAO as far as I can see
- GTD completely removed, and replaced with Tasks system that is not quite functional, especially on self-hosted.
- Scheduled tasks (not to be confused with the previous Scheduled Tasks [lobe-cron?]) don't execute on self-hosted, despite it being a full package on a server, that can surely run crons.
- Tasks is not yet stable, despite it already replacing the actually functioning GTD. Tasks don't list properly like GTD did with a nice interface in agent chat. Tasks list in top-left of agent screen doesn't update itself, requiring agent screen refresh to show in Tasks panel in Agent screen
- Agent model configuration, specifically reasoning, deep thinking, token constraints, token reasoning effort, etc, is now missing. Seems to be a known issue, but release should've been rolled back.
- Documentation is wildly outdated and does not reflect the actual features, or system. (ex. Scheduled Tasks)
- Agent chat's skills Interface adds additional steps, and doesn't allow a skill to be disabled, only auto/not. What happened? Was there a UI/UX thinking about this, or just some agent went to town?

It feels like an agent has been promoted to Product Manager as the releases are rapid, but increasingly buggy with real breaking changes. This release was the most breaking change, eliminating GTD in full, model configuration panel, and scheduled tasks / CAO doesn't actually work.

You mention Agent Bridge, to bring my own agents, like OpenClaw, Hermes, etc. I don't see such a thing in self-hosted. I only see options to add Claude Code or Codex. Agent-bridge as your screenshots above show would be great!

It would also be nice, likely available but not exposed, to allow your agents to have their own API endpoint as there's a Developer Mode and API keys, but there's no documentation on how I might be able to make specific calls to specific agents, which is a common thing. The only alternative I can see at the moment is via Channels, but you don't support all channels at the moment. So at least an API endpoint to a LobeHub agent would be a nice option. Since the whole point is to have memory and persistent agents, of course I'd want my spec'd out agent to be able to respond to requests via API so it can be integrated into workflows such as N8N, and other related apps, flows, etc.

Maybe you feature flag the items, and self-hosted is missing this core functionality? But in that case, it should NOT have been updated until it was stable and confirmed. Anyone that updated was welcomed with a relatively broken LobeHub server.

I am strong advocate for LobeHub, and promote it every chance I get. I think it's has the potential to be the best harness/interface into agents, agent groups (teams), as it truly is the most intuitive and easy to get started, and continuously refine (especially as compared with equivalent OpenClaw, Hermes, etc). However, these recent updates have made it super brittle.

I'm hopeful you can clarify a bit as the recent updates promise more than I can see delivered, and worse, 2.2.0 broke it for me completely, wasting tokens in the wind.

I can go and and on and on about ideas for features requests, and other issues (especially memory compression), but I feel that's best left after the base is actually solid and useable again.

I'm hopeful the LobeHub team will get things back on track, and reduce/eliminate these wild breaking changes.

Thank you for reading all this. Please take it all positively, as I only want good things for LobeHub!

the IM Channel is the killer feature for me. Upvoted.

How does CAO handle parallel task conflicts? Say two agents need the same resource.

How does CAO decide which skills to assemble for a novel goal? Pre trained or learns over time?

Can i manually override the agent team CAO assembles, or is it fully autonomous?

Heterogeneous agents was the technical bet I was most nervous about. Claude Code, Codex, OpenClaw — none of them were designed to be managed by something else. Took longer than we planned. Worth it.

A new skill with zero history can't compete with one that's been routed 10K times. Wonder how do you avoid rich-get-richer if that makes sense? All in all, solid work!

Hey Product Hunt 👋 Arvin here, founder of LobeHub.

Quick question before I pitch anything: how many AI tabs do you have open right now?

Claude Code in one window. Codex in another. Maybe OpenClaw or Hermes pinging you in Slack. On paper, you have an AI team. In practice, you became its operator — manually switching contexts, syncing progress across terminals, queuing up a "complex enough" task before bed because letting Claude Code idle feels like burning money.

BCG calls this "AI Brain Fry" — cognitive overload, fragmented attention, decision fatigue. 14% of heavy AI users already report it. We were promised AI would make work lighter. Somehow it made us tired in a new way.

We don't think the answer is a smarter agent. We think you shouldn't be the operator at all.

A company with a CEO but no COO is one where the founder personally chases every deadline and debugs every fire. That's exactly what your AI workflow looks like today.

So we're naming the role: CAO — Chief Agent Operator. And we're building LobeHub to be yours.

Why "CAO" and not "AI agent platform"? Because "agent tools" implies you have one agent and your job is to use it. The reality in 2026 is that you already have several agents running. This category doesn't need a better single agent — it needs a layer above them. Someone (something) to run the team.

Why this is possible now, and wasn't 2 years ago — three things shifted at once:

1. Agent self-evolution moved from papers to products. OpenClaw and Hermes proved agents can learn from sessions and turn successful workflows into reusable skills. LobeHub covers their capabilities — and goes further, because we're cloud-native: memory and skills evolve across sessions, devices, and teams.

2. MCP and Skills became the de facto standard. The LobeHub Marketplace now hosts 57k MCP servers and 270k skills. Your CAO has enough tools to actually do the job.

3. Multi-agent left the demo stage. The future isn't a single super-agent. It's an organization of agents — and organizations need an operator.

What you can do with LobeHub today:

• 🧠 Run multi-agent teams with shared memory and skills, not isolated chat windows

• 🔌 Plug into 57k MCP servers and 270k community skills out of the box

• 📡 Deploy your CAO across Discord, Telegram, Slack, Lark, and iMessage WhatsApp soon— one agent team, every surface

• 🛠️ Open source, self-hostable, and built on a runtime we've shipped to production for 3 years

I treated agents as first-class citizens on day one of LobeChat, back when "agent" still meant "a prompt with a name." Three years later, tools, MCP, skills, memory, and runtime finally compose into something that feels qualitatively different.

We're nowhere near the CAO I have in my head. Heterogeneous agent adoption, team workspaces, Agent Group 2.0 — all on the roadmap. But the direction is clear: free people from babysitting their AI, so they can spend that energy on what actually matters.

I'll be here all day answering questions. Brutal feedback especially welcome — tell me what's missing, what's broken, or what you'd want your CAO to handle first. 🙏

— Arvin, founder @ LobeHub

Been waiting months to post about this one. CAO is the update I've been quietly demoing to friends since the alpha — reactions ranged from "wait, that's it?" to "wait, that's it." Both meant in a good way. Go try it.