What's New

Main Event

• Execution Targets - We now treat laptops, SSH servers, VMs, containers, devboxes, and future hosted runners as the same concept: an execution target. Users can define local repository paths ("resources") for each execution target so that Overlord automatically launches agents in the correct place regardless of device. Each execution target simply needs the CLI installed and authenticated.
• Execution location switching — The ability to change the current target execution directory directly from the web and desktop interfaces.
• Project linking command — Overlord automatically saves a device fingerprint when an Overlord operation first runs on that execution target (device, VM, etc.), but users can now add a device directly using a new CLI command to interactively link your current working directory to a project.

Also Added

• Native browser notifications — Web-only users now receive OS-level notifications for agent alerts and updates.
• Custom launch commands (Beta) — Users can now prepend additional commands to agent launch commands, allowing integration with tools like custom execution containers (claude-pod) and other kinds of wrappers (ollama). You can find these in the "Local agent configuration" section of the CLI & Local Agents page in Settings.

Improved

• Unified notification system — Agent alerts and audio cues are now consistently delivered across desktop, web, and mobile platforms.
• Automatic CLI session persistence — The CLI now caches session keys locally, eliminating the need to manually pass session flags for every command.
• Streamlined project discovery — Standardized how projects are identified across all interfaces for more reliable tool performance.
• Flexible attachments — Relaxed strict requirements for linking files to objectives when the context can be automatically determined.

Fixed

• Real-time board sync — Fixed issues where the ticket board would fail to update automatically as work progressed.
• Duplicate session closures — Corrected a logic error that caused agent sessions to be terminated multiple times.

Added

• The Agent Runner - the desktop app used to directly launch the terminal, navigate to the working directory, and call the agent. The problem with this was poor reliability when trying to launch agents via some background process. We now use an agent runner, which automatically claims and runs queued tasks in your local terminal using a new execution queue. Clicking the "Run" button queues the objective, then the runner opens the terminal and runs it (not, you may notice a bit of delay before the terminal opens). This architecture facilitates:
• Auto-advance Objectives - Objectives can now launch automatically. Draft objectives can be set to run after the preceding objective completes. Users can now create future objectives, which automatically move to the draft position once the draft moves into execution. Each new draft then executes in sequence.
• Remote Execution- The new architecture means that the CLI-only version of Overlord can now launch objectives in the terminal as well. Future updates will leverage this to allow users to trigger execution remotely.

Improved

• Per-objective agent selection — choosing an agent for a specific task now stays scoped to that objective and no longer overrides your global default preferences.
• Optimized agent context — internal plugin documentation now uses progressive disclosure to save space and improve agent performance.
• Redesigned changelog — a new card-based layout for easier reading of product updates on the web and in-app.
• Enhanced session tracking — improved how agent sessions are linked to objectives to ensure better data integrity and more reliable task history.

Fixed

• Precise mention editing — backspacing an @mention now removes one character at a time instead of accidentally deleting the entire reference.
• Reliable tool discovery — fixed issues where the Desktop app could not find required CLI binaries on certain system configurations.
• Improved light mode readability — resolved visibility issues on public pages where hardcoded styles caused unreadable text in light mode.

Added

• Antigravity Plugins!! (beta - VERY BUGGY) - We replaced our Gemini connector, which never worked well, with a new Antigravity 2.0 connector.
• "See in Finder" button — quickly open resource directories in your native file explorer directly from the desktop app.
• Public changelog — a new public page to track product updates and improvements.
• Codex permission notifications — receive alerts for permission requests when using Codex, bringing it to parity with other AI providers.

Improved

• Queue management — future objectives now automatically advance to the draft slot whenever it becomes vacant, keeping your workflow moving without manual intervention.

Fixed

• Terminal launching — resolved an issue where auto-advanced objectives would occasionally fail to trigger the terminal window in the desktop app.
• Objective card stability — fixed a bug where clicking toolbar buttons like 'Promote' would cause objective cards to collapse unexpectedly.

Added

• Bulk add Objectives - Agents can now use the CLI and MCP to add a sequence of objectives to a ticket. (rather than just adding a bunch of tickets)
• Public Changelog — Now you kids get to see what I changed.

Improved

• Objective auto-advance UX improvements - auto-advance now defaults to false, and the toggle for activating it per objective should now be clearer.

Fixed

• Fix Codex UserPromptSubmit hook — Codex agents should now more reliably record user response in chat.