Changelog

Hermes Agent can now use custom functionality from plugins for every connected platform, not just Telegram. This allows plugins to add new features and integrations for any platform Hermes supports without needing changes to the core code.

The AgentMail email skill has been updated to prioritize command-line usage, making it easier to interact with and configure. This change ensures that users can still set up the skill without needing an API key upfront, supporting a simpler self-signup process.

A new optional tool can now help you publish your websites online to services like GitHub, Cloudflare, or Netlify. This makes it easier to put your projects on the internet and manage different versions of your site.

Previously, setting a very large timeout for approvals could cause Hermes Agent to crash on macOS. This has been fixed by ensuring that the timeout value is kept within safe limits, preventing the crash and ensuring smoother operation.

The gateway's watchdog can no longer mistakenly shut down a healthy gateway due to a slow internal check. It now requires two separate signals to confirm a problem before taking action, preventing accidental service interruptions.

This update improves how Hermes handles crashes during data saving by making sure it accurately tests different saving methods. It also fixes how it detects and reports errors, making it more reliable for ensuring data isn't lost.

Cron jobs that timed out will now properly stop all their child processes, even those that started in their own session, preventing them from running unexpectedly in the background. This ensures that timed-out jobs are fully terminated, improving system stability.

The Hermes Agent now explicitly sets a timeout for when it checks with a smart approval system. This prevents the agent from getting stuck for a long time without any feedback if the approval system is slow or unresponsive.

Plugin code that runs during important agent steps can now be stopped if it takes too long, preventing it from freezing the entire conversation. This ensures the agent remains responsive even if a plugin gets stuck.

The dashboard can now properly recognize and trust specific reverse proxies when connecting, which is important for secure authentication. This change ensures that login processes work correctly even when the dashboard is running behind a proxy.

When a browser command takes too long, Hermes now better handles the situation by either resetting the browser session or, if the browser process itself is unresponsive, forcefully shutting it down to ensure stability. This prevents the agent from getting stuck or returning incorrect information due to a hung browser.

The desktop application can now start a more lightweight server for communication, improving its startup speed and reducing its complexity. This change is optional and can be enabled for a more streamlined experience.

A bug where Hermes incorrectly thought subprocesses were dead has been fixed. This means commands that rely on these subprocesses will now run correctly instead of failing unexpectedly.

Hermes Agent can now keep information like variables and loaded data between code execution steps for more complex tasks, improving efficiency. This new "session" mode makes it easier for the agent to remember context without needing to reload data repeatedly.

The system now correctly identifies and cleans up old sessions that are truly no longer in use, preventing active sessions from being accidentally deleted when multiple Hermes processes share the same data. This ensures a more stable and reliable user experience.

You can now control whether links automatically show previews in Slack messages sent by Hermes. This feature works for both direct message sending and when Hermes is relaying messages for other services.

Voice messages can now be transcribed again by correctly sending media information and specifying the message type. This also fixes issues with other types of relayed media from Discord, such as images and documents.

Hermes can now use your logged-in browser sessions to browse the web. This allows the agent to act as you, using your cookies and sessions, without opening your live browser or breaking on newer Chrome versions.

Previously, Hermes Agent could silently shorten precise decimal numbers when saving them to configuration. This change now ensures that very precise decimal values are kept exactly as entered, preventing data loss for important identifiers.

Updates on Windows desktop will no longer hang indefinitely or incorrectly report as failed. This change ensures the update process completes successfully and is accurately reflected.

Messages you see instantly are now correctly separated by profile, preventing them from appearing in the wrong conversation. This ensures a cleaner and more reliable chat experience.

The way Hermes Agent describes and handles shell commands has been improved for accuracy and simplicity. This change ensures commands are written for your specific operating system and includes a new, unified way to get notifications when commands finish or meet certain conditions, making it easier to use and understand.

The Hermes Desktop can now manage the complete update process for remote SSH instances. This allows you to update your fleet of remote devices directly from your desktop, ensuring they stay up-to-date and secure.

Hermes Agent now correctly applies configuration updates even if an initial update attempt fails and needs to be retried. This ensures you always get the latest settings and important prompts for your Hermes Agent.

Helper processes that handle communication have been improved to ensure they are properly cleaned up if their main process crashes. This prevents lingering processes that could consume resources.

The project's code formatting has been automatically improved, making it easier to read and maintain. This update ensures consistency and helps prevent common coding errors.

The desktop updater now waits 10 minutes before assuming a step has stalled, preventing it from interrupting legitimate long processes. This change makes the automatic update process more robust for users running slower builds.

This update improves the stability and reliability of Hermes Agent's connections to external services by fixing several issues related to how it manages these connections. It ensures that the agent correctly handles connection errors and keeps local settings consistent when switching between different connection types.

Several important backend improvements have been re-applied to ensure the gateway cluster is more stable and responsive. This includes better handling of configuration changes, improved search functionality, and fixes for potential system slowdowns.

You can now manage updates for individual SSH connections directly from the Gateways settings. This new section shows the update progress for each connection and displays a detailed receipt of what happened during the update.

The Bots picker now reliably settles on a selection and avoids unnecessary refreshes, preventing data loss during edits. Additionally, changes ensure that when you refresh the model list, your chosen model is kept if it's still available, or a new one is automatically selected, preventing unexpected errors. These updates also mean that important confirmation prompts for sensitive model changes are now consistently shown across the application, protecting you from accidental configurations.

When a bot chat loads slowly, it now has more time to connect and start working properly. This prevents it from appearing broken just because the backend takes a little longer to prepare.

Hermes Agent now correctly stops if a profile's storage can't be opened, preventing data from being mixed with the wrong profile. This ensures profile data stays organized and prevents unexpected behavior when profile storage is temporarily inaccessible.

The project's code formatting has been automatically updated to ensure consistency and improve readability. This change helps maintain a clean and organized codebase for all contributors.

The project's code formatting is now automatically checked and corrected using `npm run fix`. This ensures a consistent and cleaner codebase for everyone.

Managed SSH connections can now be reliably established on standard Ubuntu systems, even with complex file paths. This fixes issues where connections would get stuck or fail due to incorrect handling of quoted paths and differences in shell commands.

The desktop interface now shows more descriptive titles for browser actions, making it easier to understand what the agent is doing. Previously, these actions were all labeled generically, but now they are titled based on the specific instruction comment, just like in the command-line and text interfaces.

When the Hermes Agent is stopped unexpectedly, it now saves unsaved chat sessions to prevent data loss. Additionally, changing compression settings now correctly reverts to default values if a setting is removed from the configuration.

Existing chats that were previously unopenable after updates will now be accessible again. This allows you to view your chat history even if the owner information was missing for older sessions.

The option to use a WebSocket-only connection for the desktop interface has been removed. This simplifies the desktop backend and ensures that a recent fix for event replay is correctly applied.

When the Hermes Agent restarts a session after it was interrupted, it now correctly remembers who was in control of that session. This prevents issues where the agent might get confused and stop responding to commands.

Chat platforms will no longer receive duplicate or unwanted "Context compaction complete" messages. These notifications now only appear when compaction actually succeeds, making them less intrusive for users.

The Hermes Agent will now refuse to overwrite your configuration file if it's unreadable or improperly formatted, instead creating a backup of the corrupted file. This prevents accidental data loss when modifying settings.

The way Hermes Agent reconnects to cloud services when interrupted has been improved. It now uses the correct sign-in process for cloud accounts, making it easier to get back online.

The Hermes Agent can now correctly find its default configuration directory on Windows, ensuring tests for this functionality pass on that platform. This fix makes the project more robust and reliable for all users, regardless of their operating system.

The cron tool will now more accurately detect if the gateway is running, preventing false warnings when jobs are actually scheduled to run. This improves reliability by prioritizing a runtime lock to confirm the gateway's status.

The Hermes Agent can now find and use the free minimax/minimax-m3 model from OpenRouter. This makes a powerful free AI model available for users who prefer using OpenRouter for their AI needs.

Group chats no longer show empty messages or hide real answers from you. This makes sure you see all the responses from group conversations.

The way bots take turns when mentioned has been improved so that a bot that is @mentioned will always get its chance to respond, even if there are limits on how many turns bots can take. This also makes it clearer when a conversation ends because of these limits versus when everyone has agreed.

Group chats now have a "Stop" button that lets you immediately end the current turn and pause the rest of the conversation. This makes it much easier to control ongoing group discussions.

The project's code formatting and linting have been automatically improved to ensure consistency and catch potential errors. This helps maintain a cleaner and more reliable codebase.

New free versions of Inkling and MiniMax M3 models are now available in the OpenRouter selection. This makes these powerful, cost-effective options accessible to more users.

The GLM-5.3-Flash model can now be easily selected in various places, including the setup wizard and code generation tools. This makes it more accessible for users to leverage its advanced capabilities.

The `openrouter/elephant-alpha` model has been removed because it's no longer available from its provider. This update ensures that Hermes Agent only offers working models, preventing users from selecting unavailable options.

Hermes Agent now correctly allows you to use free OpenRouter models even when only free models are allowed. This prevents unnecessary errors and ensures providers are only marked as unhealthy when there's a genuine connection issue.

The Hermes Agent now uses fewer words to describe how it interacts with the browser, making it more efficient without sacrificing accuracy. This change helps the agent perform its tasks more quickly by streamlining its internal instructions.

Hermes now correctly sends important startup messages to the correct output stream, preventing the desktop application from timing out while waiting for the backend to start. This ensures a smoother startup experience for users.

This update prevents duplicate message replies when a platform's server is slow and causes a timeout. Previously, Hermes might send the same message twice if a timeout occurred after the initial delivery.

This change removes a duplicate class that was confusingly named, preventing potential future errors when using the deadline feature. Now, when referring to `SuspectableBackend`, you'll always get the correct version.

Hermes Agent has been updated to version 0.20.6, bringing together a collection of recent improvements. This release ensures the agent is stable and ready for continued use with minor fixes and a version update.

Hermes can now correctly launch the CUA Driver on macOS, even when installed via a standard shortcut. This fixes an issue where the driver might not have been recognized due to changes in its security signature.

Dictation from Mistral Voxtral will now correctly show your spoken words instead of the raw technical data. This update ensures a smoother and more user-friendly experience when using this speech-to-text feature.

The Hermes Agent now properly closes database connections when errors occur, preventing data leaks. It also avoids opening unnecessary database files when tasks are skipped, improving efficiency.

The system now correctly preserves archived conversation data even if a session appears empty, preventing accidental data loss. This ensures that past conversations remain available for review after certain operations.

The way Hermes decides which tasks to delegate and how it structures those tasks has been simplified, resulting in a 36% reduction in the token count per call. This change makes the interface clearer and more efficient for the agent to manage its tasks.

The Hermes Agent can now send messages to WeCom more like a real conversation. You'll see a typing indicator, and replies will appear word by word as they are generated, making the interaction feel smoother and more responsive. This also fixes issues where messages might have been sent twice.

The gateway now correctly keeps your chat session information after a crash, preventing lost progress. Previously, a crash could cause the system to incorrectly remove your session data, requiring you to restart your conversation.

When a session disconnects and then immediately reconnects, it will now correctly stay connected instead of being shut down. This prevents data loss or interruptions for users who experience temporary network glitches.

Hermes Agent now informs you if it had to ignore your database settings, and you can reliably control how often your database saves changes across all platforms to prevent data loss. This makes it clearer when your chosen settings aren't being used and improves data safety.

The update process now correctly prevents accidental deletion of your personal files that are ignored by Git. This ensures your local data remains safe even when Hermes needs to use a fallback method to update your project.

The system's shutdown process now correctly accounts for all potential delays, preventing premature restarts that could interrupt tasks. This ensures a more stable and reliable shutdown experience.

Manual triggers for recurring scheduled tasks will now correctly execute. Previously, a bug prevented these manual runs from actually starting the job.

The system now correctly marks runs that fail unexpectedly, like after a tool use or during an API wait, as incomplete instead of incorrectly showing them as completed. This prevents invisible errors in run history and ensures more accurate reporting of run statuses.

When a connection to an MCP server fails unexpectedly, Hermes Agent will now attempt to restart the server automatically. This prevents the server from staying broken and unresponsive for a long time after a single failure.

Hermes can now send images from tool results to Gemini 3.x models, allowing them to see and process visual information. This fixes a bug where images were previously dropped, preventing Gemini from using them.

Provider names that were accidentally read as numbers are now correctly treated as text. This fixes an issue where custom AI model providers might not appear or be usable in the Desktop application.

The project's automatic formatting tool can now fix code style issues by running a simple command. This helps keep the code clean and consistent without manual effort.

This update ensures that when two parts of Hermes Agent are using the same profile, one won't accidentally shut down the other's connection if it loses its own. This prevents interruptions to ongoing tasks and data saving, even if one connection drops.

Messages that were previously hidden at the top of a long chat are now correctly displayed when you click "Show earlier messages." This ensures that the full conversation history is accessible, even in lengthy sessions.

Hermes Agent now ensures only one part of the system is in charge of setting the database's journal mode. This prevents accidental data corruption and ensures that critical operations like repair and delegation work reliably.

We've updated our system to correctly recognize two contributors for their work on compression improvements. This ensures their contributions are properly credited and allows the release process to continue smoothly.

Clicking on a bot in the Bots list will now correctly keep that bot's chat window open and focused, preventing it from unexpectedly closing or sending messages to the wrong conversation. This change ensures a smoother and more predictable user experience when interacting with bots.

The Hermes Agent desktop app now ensures that collapsed panes, like the Cronjobs tile, always leave a visible way to restore them, preventing them from vanishing completely. This change also makes sure the Sessions and Bots strip can't be accidentally hidden.

Hermes Agent can now send messages properly when you're using remote connections like Docker or SSH. This fix ensures message deliveries always run on your main machine, preventing errors that occurred when trying to run them inside a container.

When the compression process gets stuck, Hermes Agent will now retry it once using a fallback option before giving up. This helps ensure conversations are compressed more reliably.

Previously, when the conversation history was being compressed mid-conversation, the same user message could be saved twice. This update ensures that each message is only saved once by using a more reliable method for tracking what has already been saved.

Recurring jobs that get stuck in an error state can now be resumed instead of needing to be deleted and recreated. This ensures that scheduled tasks will not be permanently disabled due to temporary errors.

Labels for older skills in the journey view will now remain readable against dark backgrounds. This change ensures that important information is not lost due to low contrast when displaying historical data.