Smart Node moved from release 1.12.23 on 1 September 2026 to 1.12.48 on 29 September. Much of the work in that time addressed how a node handles daily operations on an existing server: running updates safely, preventing agents from talking over one another, and enforcing security boundaries at the system level.
Here is what changed during the month and what it means for your office.
Chat: attachments, continuous typing, and visibility
The office chat received several practical updates for day-to-day work:
- File attachments: You can drop files onto the chat field, choose them from your device, or paste a screenshot directly. Agents can read text files, PDFs, and images. Attached files stay on your node alongside their conversation and are removed if you delete the chat.
- Non-blocking turns: You do not have to wait for an agent to finish working before typing another instruction. You can continue sending messages, and each one receives its own response. A running reply now has a Stop button.
- Visible progress: Instead of an uninformative waiting icon, the chat displays each step an agent takes and what it outputs while it works. A Console view on each agent card provides this same breakdown for actions triggered from any channel.
- Session capacity limits: A meter shows how full the active chat session is. When the session reaches its limit, the agent saves a handoff note to memory and starts a fresh session, rather than quietly compressing earlier messages.
- Mobile and approvals: On a mobile browser, the input bar stays visible above the keyboard and toolbars, and the sidebar opens as a full-screen view. The Approve, Modify, and Reject controls work in the office interface, with Modify and Reject prompting you for an explanatory note.
- Sign-in duration: Sessions expire after 12 hours of inactivity or 30 days from sign-in.
Agent collaboration and routing
September updates corrected several issues with how agents coordinate in the background:
- Returned answers: When an agent consults a colleague, that colleague's response now routes back to the specific office chat where the question originated. Background tasks triggered by a colleague end with an answer sent back to that colleague.
- External messaging: On channels like Telegram, agents no longer send a second summary message after delivering their answer.
- Turn separation: Agents run a single turn at a time. A scheduled wake-up and an incoming message can no longer execute the same agent simultaneously.
- Chat isolation: Opening two separate office chats with the same agent maintains two distinct conversations. The agent will not mix them up or answer a question from the other thread.
- Internal email: Agents can communicate with each other over email, and messages sent to multiple node addresses now reach every addressed inbox.
Security checks handled by the node
Several safeguards that previously relied on written prompts are now built into the node runtime:
- Access to secrets: An agent cannot read a stored secret without an explicit grant. Writing and deleting secrets is restricted to the owner and agents specifically designated by the owner.
- Self-assigned tasks: If an agent creates a task for itself, that task requires owner approval before it can wake any agent. The node also caps the number of open self-assigned tasks an agent can hold.
- Data protection: Outbound messages containing requests for credentials are blocked from leaving the server. Unsolicited email from unapproved senders is excluded from an agent's context window.
- Process isolation: Agent instructions no longer show up in the host process list, and model credentials do not get written to system logs. In the office interface, documents such as HTML and SVG open within a sandboxed environment that blocks scripts.
Runtimes beyond Claude
Smart Node now supports models other than Claude. The Gemini runtime arrived in 1.12.42, and other external models can connect through OpenRouter.
The configuration interface groups available models by runtime and displays their full names. The setup agent can assign an agent to a new model directly without requiring manual database edits, and it will notify the owner if a chosen provider is missing an API key.
Smaller updates and updater fixes
For anyone maintaining a server, the update process is smaller and more transparent:
- Smaller download size: Memory, task queues, and voice processing were moved from separate containers into standard system services that use the existing node runtime. Release packages dropped from around 950 MB to roughly 300 MB. Updating one of these three services requires a service restart rather than a full release download.
- Fixes on existing nodes: Releases are now tested against servers with existing data and history, rather than only clean machines. This addressed upgrade bugs involving settings an update reset, files left behind by older releases, database steps colliding with running agents, and checks that ran the moment a service restarted.
- Accurate update reporting: The updater now waits for the installer to finish and reads its status. An install that provisions zero agents reports a failure rather than a false success. Releases 1.12.30 and 1.12.31 were pulled prior to public release when they failed on an active node, and were superseded by 1.12.32.
- Explicit testing notes: Changelog entries document gaps when something could not be verified. For instance, the voice line in 1.12.30 was released as code without a claim that it worked, because a test call failed.
Full notes
To see the technical details, reasons for each fix, and testing records for every version from 1.12.23 through 1.12.48, read the September changelog. To see the latest releases, visit the general changelog.