Wispr Flow: Dictation That Works EverywhereStop typing. Start speaking. 4x faster.
xAgent
Self-hosted AI agents for real work
10 followers
Self-hosted AI agents for real work
10 followers
xAgent is a self-hosted, multi-user AI agent portal for teams that need work completed, not just chats. Run it on your own server, prepare shared Skills, Tools, MCP services, connectors, approval policies, and isolated workspaces, then let users submit tasks through the web UI or IM. The free beta supports up to five users and keeps all data in a fully private deployment. It includes WeChat and Telegram connectors.
How does the approval policy actually work in practice — do users get blocked from running tasks until an admin signs off, or can they queue things up and approvals happen asynchronously?
@tugay865925 Approvals are applied to specific actions during execution, not as a blanket gate before a user can start every task.
xAgent has two layers of policy:
Public approval policy: configured by the administrator as the system-wide safety baseline. It defines which actions can run directly, which require approval, and which are denied. Users cannot weaken this baseline.
Personal approval policy: an optional, stricter layer that advanced users can tune for their own work. It can add more approval requirements, but cannot bypass the public policy.
In practice, a user can submit and start a task normally. If the agent reaches an action that requires approval, that action waits for approval before it runs; the rest of the task does not bypass it. Actions allowed by policy continue automatically, while denied actions are blocked.
This lets teams reduce operational risk without forcing an administrator to manually approve every task or every AI interaction.
@tugay865925 One practical observation: many individual users eventually disable approvals because too many prompts become disruptive. That can be acceptable for personal use, where the user accepts the consequences, but it is usually not acceptable for a company environment.
The current approval system is still an early implementation. The next step is to unify the approval architecture so approval requests can also be delivered through IM connectors. Users will be able to review the request and approve or reject it directly from channels such as WeChat or Telegram, with an optional comment.
The goal is to keep enterprise safety controls in place without turning approval into a workflow people simply bypass.
set it up over the weekend and the approval policies plus isolated workspaces actually made it usable for my small team, not just another chat wrapper. nice touch including the wechat connector out of the box.
@demirel_ye77608 I finished the Telegram connector over the weekend; you can give it a try too.
Love that the approval policies and isolated workspaces are baked in from the start—makes it way easier to hand this off to teammates without worrying about one person's task stomping on another's data.
@necdet644676 Yes, I place a strong emphasis on security. That’s why every user has an isolated workspace—though users don't have 100% control over that space, given that most simply care about functionality and aren't overly concerned with privacy or security. As for the built-in approval process, it essentially allows the security-savvy members of our system team to establish a unified line of defense for everyone.
The self-hosted setup with shared Skills and approval policies feels like exactly what small teams need right now. Surprised how clean the task submission flow is, and the WeChat connector actually just works without fuss.
@masalkavut I finished the Telegram connector over the weekend; you can give it a try too.
# xAgent 0.0.5.beta Is Available
xAgent is a server-side, multi-user AI Agent work portal built for completing real tasks, not just conversations.
Administrators prepare models, Skills, Tools, MCP servers, Connectors, and safety policies. Users then describe the goal, provide materials, and confirm important actions when needed. Tasks continue on the server even when the user's device is offline.
This release focuses on runtime and service-boundary consolidation rather than adding more surface-level features:
- Clarified responsibilities across runtime flows to reduce maintenance and troubleshooting complexity;
- Prepared the codebase for progressively consolidating Connectors, installation, and release workflows in the main repository;
- Continued strengthening existing session, approval, Skill, Tool, MCP, Connector, and workspace-isolation capabilities.
The goal is to make xAgent a tool that gets work done. It can chat, but when a task involves files, external systems, approvals, or long-running execution, it should keep moving the work forward.
xAgent is still in beta. It is intended for people exploring self-hosted deployment, multi-user AI Agents, custom Skills, and IM Connector integrations.
Documentation: <https://xagent.xiagaogao.com/en/>
Feedback is welcome:
1. Which recurring tasks would you most like an Agent to run continuously?
2. Where do you get stuck most often: deployment, model configuration, or Connector setup?
3. Which Skills, MCP integrations, or Connectors would be most useful?
4. Report an issue or share a feature request:
<https://github.com/coffeehc/xagent-releases/issues/new>
the self-hosted setup was painless and the approval policies actually make sense for letting non-tech teammates run agents safely. five free users is generous for kicking the tires.
@nihatho6p For families or small teams, a 5-user limit should be sufficient. I hope everyone can effectively leverage AI to assist with their work; after all, I believe that technology shouldn't just be about making money—it should also be driven by a desire to change the world.