Changelog

• Gitea Support — Bring agents into Gitea issues and pull requests, with automated reviews and commit status updates.
• Improved Webchat — Use interactive MCP Apps directly in webchat.
• Broader Sandbox Protection — API key protection now covers more runtimes in MicroVM sandboxes, including Claude, Codex, and OpenCode.

Behavior change: The Console's "Run again" action for code-host sessions has been removed. Use a follow-up message in the session to request another run.

Upgrade note: Start a new OpenCode session to enable its added API key protection.

• Centralized Memory Management — Keep managed memory available across machines and while sandboxes are idle. Browse and edit managed or external memory through a shared Console interface.
• Improved Chat — Answer agent forms in Discord, Feishu, and Telegram, and continue webchat after approving an operation.

This release also improves skill-install recovery and code-review delivery after connection interruptions.

Behavior change: Managed-pool agents migrate managed memory to the Control Plane. Returning that memory to a daemon requires a forced reset and does not copy the existing memory back.

• MicroVM Sandboxes — Run agents in opt-in MicroVMs on Linux hosts with KVM, with separate VMs for isolated sessions and persistent workspace disks.
• GitHub Deployment Triggers — Let agents react to deployments and status changes, keeping context across deployments to the same environment.
• Improved Chat — Answer richer agent questions, including typed responses and forms, across webchat, Slack, and Telegram.

Upgrade note: Custom sandbox read/write root settings must be converted to sandbox.mounts. GitHub Apps need deployment permissions and event subscriptions to use deployment triggers.

• Improved Webchat — Answer agent questions with interactive choices and multi-select cards, and find agents more easily in the composer.

Follow-up code reviews now reuse installed dependencies and built workspace files, and Slack final replies are delivered more reliably.

• Session Isolation — Isolated sessions get separate workspaces and runtime homes, with separate sandbox pods on Kubernetes.
• Webhook Conversations — Group webhook events by subject so updates to the same ticket, order, or conversation continue an existing session.

Upgrade note: Update the Control Plane and relay together before enabling subject-based webhook sessions.

• Linear Support — Delegate issues to agents, follow their activity in Linear, and connect their work back to pull requests.
• Approval Notifications — Receive approval requests through your linked chat accounts and the Console notification center.
• Improved Webchat — Continue webhook-origin sessions and follow live agent plans from the Console.
• OpenClaw Support — Run agents through an installed OpenClaw gateway; execution remains outside AgentConnect's sandbox.

Upgrade note: Existing Slack apps missing the expanded permissions now show a reinstall notice. Reauthorize them to enable the affected tools.

• Git URL Workspaces — Connect repositories by Git URL when this option is enabled for your installation.
• Improved Slack — Follow native plan and tool progress, stop the correct running turn, and see clearer session titles.
• Improved Webchat — Reconnect to running conversations after a page reload.

New schedules default to your local timezone.

Subscription change: New pull-request and merge-request subscriptions default to every update; existing subscriptions keep their rules.

Behavior change: Workspace and skill-source cloning now allows any valid Git origin by default. Configure an explicit origin list to restrict hosts.

Connect a new daemon directly while editing an agent's placement, then select a runtime and model supported by that machine. This release also fixes Claude collaboration-tool routing and stale sandbox-startup notices.

Configure separate subscriptions for issues, pull or merge requests, and pushes in the same repository, each with its own trigger rules. This release also improves skill installation for DeepSeek Harness and Kubernetes sandboxes.

Agents running in Kubernetes sandboxes can now install configured skills, managed skills, and approved Dream skills before their runtime starts.