Hey everyone, Nathan here from CopilotKit
We re getting ready to launch OpenBot, a self-hostable workspace for AI agents with built-in governance.
You can bring agents your team has already built through AG-UI, connect tools through MCP, and choose your models and infrastructure.
For business teams, that means control over how agents work:
Hey Product Hunt, Nathan here from CopilotKit 👋
Before an AI agent can access company systems, someone has to be accountable for what it can read, change, send, and spend.
That’s the focus of OpenBot: a workspace for AI agents with enterprise governance built into how they operate.
Your company decides who gets access and what agents are allowed to do.
• Identity and permissions: SSO and SCIM connect access to your company directory. Bots can act on a person’s behalf with scoped permissions. Tool access can be set per Bot, role, and person.
• Approvals and restrictions: Decide which actions can run, which need a person’s approval, and which are blocked. An agent might be allowed to research an account but need approval before emailing a customer.
• Auditability: Track the information agents access and the actions they take. Export audit events to your existing SIEM so your security team can investigate activity in the systems they already use.
• Spending controls: See and control usage across Bots, teams, roles, and people.
• Infrastructure ownership: Self-host on your own infrastructure, with control over tenancy, data retention, and deletion.
You also keep the stack your engineers chose.
Bring existing agents through AG-UI, connect tools through MCP, and choose your models. You can customize or white-label the workspace for your organization.
Agents can work with company knowledge, research, analytics, and operational tools. They can also learn processes and develop reusable skills while staying within the permissions and approval policies you set.
That last part matters: an agent getting better at a task shouldn’t mean it gets broader access.
OpenBot is available for Mac and Windows, free for personal use. OpenBot for Business is in early access.
Explore it here: https://open.bot/
For the people responsible for deploying agents at work: what would your IT or security team need to verify before approving one? A specific workflow or requirement would be especially useful. We’d love to hear it.
Thanks for taking a look.
@nathan_tarbert Awesome! The day has come 🎉 We were waiting so much for this.
@eli_berman2, I'm so excited to get OpenBot into the hands of users! 🚀
@nathan_tarbert exciting to see OpenBot offers an enterprise solution!
Thanks, @craig_irwin! Governance is a big part of why we’re building OpenBot for Business. Companies need control over what agents can access, which actions require approval, what they spend, and where their data lives—with an audit trail to back it up.
We’re closing the gap by bringing those controls together in one self-hostable workspace, while letting teams keep the agents, models, and tools they choose.
Does it work on Claude/ChatGPT subscription?
@danshipit Yes! OpenBot Desktop lets you sign in with your Claude or ChatGPT plan when you set it up, so you don't need an API key.
You can also use an API key (OpenAI, Anthropic, Gemini), a Google or xAI account, or your own model endpoint. To switch later: Settings → General → AI model.
@eli_berman2 That's great and super important!
@danshipit @eli_berman2 does it wrap claude code cli/codex cli or use oauth directly?
@danshipit yeah you can sign in with your Claude, ChatGPT or xAI (Grok) plan. Gemini and Unbiased AI work too. and you can run local models like DeepSeek, Kimi, Qwen and Mistral with Ollama & others
@danshipit, we've tried to make it as easy as possible for you to go from zero to automating your tasks as quickly as possible. Connecting to your ChatGPT/Claude account during setup was an important step, and I would love your feedback after you try it.
You can download the desktop app at https://open.bot/
I tried Grok Bot for a month but the recent outages made it pretty frustrating.. this one fixes most of what bugs me about coworker products from big companies. it all runs on my machine, the skills it learns are just files I own and I can bring my own agent as a developer and pick a cheap/local model per bot.
I already pay for Claude Max so being able to sign in with it, letting bots hand work to each other and each bot having its own workspace & browser is insanely productive. great launch team, hope it keeps getting better.. big fan of the open source ecosystem!!
@anmol-baranwal Lets go! Since OpenBot is built on CopilotKit and AG-UI, it's architecture makes it actually enterprise-ready. For proof: majority of the Fortuna 500 already use both, in production. Many other open source products out there just have tons of slop code... We built this with real companies and businesses in mind
Thanks, Anmol! The ownership piece matters a lot to us too-your skills stay as files you own, and you choose the agents, models, and infrastructure.
I’ve used Grok Bot as well, and for business use, having control over where work runs and what data reaches a model provider matters. Self-hosting is part of that; permissions and approvals matter just as much.
Really appreciate you putting OpenBot through its paces. What work have you been handing between bots so far?
Looks promising, but I'm stuck on install. I use Orbstack, but looks like the installer requires Docker Desktop or Podman?
@madsb Thanks for giving it a try, and sorry for the snag! You’re right: the desktop installer doesn’t support OrbStack yet. It runs on Docker Desktop or Podman.
Quickest way in today: on an Apple silicon Mac, quit OrbStack and run the installer again, and OpenBot will set up its own Podman engine for you. On an Intel Mac, install Docker Desktop or Podman Desktop first.
OrbStack support is on our radar, and you can follow it here: github.com/CopilotKit/OpenBot/issues/1555 🙏
@madsb @eli_berman2 Orbstack works just fine, I ran through it a few days ago. Are you running into a particular issue with it @madsb ?
@madsb @mxmzb Oh really, that's good to know..!
Finally! I don't have to keep 2-3 subscriptions to use the best models for my day to day AI work!
@jerel_john_velarde Yep, you can bring the one(s) you already have!
@jerel_john_velarde, Thanks!
Having your model choices in one workspace is very important. You can connect your existing provider accounts or API keys and choose what works best for each task. Provider subscriptions and API usage still have their own costs but it gives you the option to choose which LLM or harness you want to use.
How easy is it to add gateways to chat apps?
@arctictony, pretty easy.
Slack and Microsoft Teams come through OpenTag, our open-source chat gateway (https://github.com/CopilotKit/OpenTag). You run OpenTag, point it at your OpenBot, and set a shared secret on both sides. That’s about three environment variables, and you don’t create a separate Slack app for OpenBot.
Each person then links their Slack or Teams account from a Bot’s page with a one-time code, and talks to that Bot from there. Approvals show up as cards right in the chat, and every turn still goes through the same permissions and audit trail as the web app. SMS works too, through Twilio.
Setup: https://github.com/CopilotKit/OpenBot/blob/main/docs/configuration.md#slack-and-microsoft-teams-opentag-pairing
@nathan_tarbert Sweet! Will check it out!
@arctictony, that sounds great. You can reach out to me directly on X @nathan_tarbert. I would love to get your feedback.
@nathan_tarbert @eli_berman2 what happens when an AI agent encounters a task outside it’s knowledge base ? Can it intelligently escalate to a human ?
@nathan_tarbert @dipanshu_kushwaha5 Yes! OpenBot comes ready with human-in-the-loop approvals. So when your bots come across sensitive tasks, they pause and wait for the user to approve/decline. Also, any user can takeover a bots work whenever they choose, and hand it back when they're ready.
@nathan_tarbert @dipanshu_kushwaha5 @eli_berman2 I really like that OpenBot offers human-in-the-loop approvals for Email Operations! For me, sensitive tasks include sending out email which cannot be undone.
One hundred percent @heidi_spitz1, human approval is built in by default, and it's really important when any destructive action can take place.
Good question @dipanshu_kushwaha5 . A task outside an agent’s knowledge and an action that needs approval are two different cases. Eli covered the approval side; that alone doesn’t mean the agent will reliably recognize every knowledge gap.
You can take over its work and hand it back once you’ve provided the missing context. I would love to dig into this. Is there a particular workflow you’d want to escalate automatically?
@dipanshu_kushwaha5 great question, I can answer this from my own experience - each bot has its own job with instructions, skills, tools and goals (the assigned responsibilities). so when something falls outside that, it doesn’t guess & passes the task to the bot that owns it, and that bot can pass it on again if needed.
if no bot fits or the action is sensitive, it brings you in.. you can approve, decline or take over. workspace works like a real team where everyone knows their role and who to ask
@nathan_tarbert @eli_berman2 @dipanshu_kushwaha5
@dipanshu_kushwaha5 adding the protocol side: pausing for a person is part of the protocol. A run can finish with an interrupt that says what the agent needs, and the next run picks up from the person's answer. So "wait for a human" doesn't depend on keeping a connection open, and any agent that speaks AG-UI can do it the same way.
@markusmxe, appreciate you adding the technical details on how it works!
@nathan_tarbert @dipanshu_kushwaha5 @markusmxe This is coming from the guy who was the mastermind behind AG-UI protocol!