August 31, 2026
Nous Research made Bot Mode a bundled, default-on part of the Hermes Agent desktop app in version 0.21.0, released on August 31, according to the release notes. Each agent profile becomes a named bot with its own model, memory, skills and chat history, stored on the user’s machine.
Takeaway points
- Group chats hold two to six bots, with hard caps of 10 messages per send and three rounds, according to the Bot Mode documentation.
- A new peer command lets agents message one another across machines; StrideNote found no cost controls or permission model in the Bot Mode documentation.
- Bot Mode ships only in the desktop app; on StrideNote’s test machine, command-line version 0.21.1 includes the peer command and an OpenClaw migration tool.
What is Bot Mode in Hermes Agent?
Nous Research calls version 0.21.0 “The Pantheon Release.” Bot Mode appears as a Bots tab beside Sessions in the desktop app’s sidebar, according to the documentation, and needs no separate install. There is no new object underneath: a bot is an ordinary Hermes profile, with its own configuration, memory, skills, credentials and chat history in its own folder, so the same agent can be opened from the command line. The documentation says Bot Mode adds “no core patches, no background daemons, no extra storage.”
Every bot gets a name, a generated avatar that can be randomized or locked, and a place in a shared roster. Each has one permanent Bot Chat, created when the bot is made; typing /new there compacts the context instead of starting a fresh session. A new bot needs a name, a title and a description, and an advanced panel lets a user clone another profile or pin the bot to any provider and model pair Hermes supports, so one bot can run on a local model while another uses a hosted API. Local model tool calling in Hermes and OpenCode covers the setup such a bot relies on.
How Hermes Agent bots talk to each other in group chats
Group chats hold between two and six bots. The documentation says “hard caps (10 messages per send, 3 rounds) keep rooms from spinning,” and that a bot speaks only when it has something new to add. Users can mention any bot by name. The roster can show only bots active now, meaning those that wrote within the last 90 seconds or report a recent worker heartbeat, and a user can hide a bot from view without stopping its routines or removing it from group chats. Bots can also be filed into sections, which travel with the bot’s profile to any desktop connected to the same backend.
Bots can also message each other directly. Inside the desktop app a message_agent tool queues the message and delivers replies as background notifications. Across machines, the new peer command registers another Hermes gateway by its API server address and key, then sends a message to a named agent there. On StrideNote’s test machine, the command’s help text lists subcommands to add, list, remove, message and check on peers, and says the remote key is stored as a credential in the local configuration.
Bots can carry routines, recurring tasks that appear in the scheduler as jobs named after the bot and post their results into that bot’s own chat.
What else is new in Hermes Agent 0.21.0
Version 0.21.0 rolled up about 5,800 commits and 2,475 merged pull requests from more than 760 contributors since 0.20.0, the release notes say. Scheduled jobs now load and update persistent memory, can carry one run’s output into the next with a continuity setting, and can skip the model entirely when a monitored source has not changed. A running subagent can be listed, given a course correction, or stopped with its partial result kept, and the defaults for delegated work rose to 250 iterations and 10 concurrent children.
The release also merged MCP server management into one desktop page with background health checks that prompt a user to re-authorize before a tool call fails, a cost and usage overlay, and hermes:// links that install an MCP server after explicit confirmation, let the agent drive the desktop app’s own browser, and added six model providers, among them the Meta Model API. The command line gained a command palette and a global emergency stop.
Four days earlier, version 0.20.6 had added opt-in browsing with a managed copy of the user’s own Chromium profile, gated on consent, and opt-in encryption of stored secrets in the operating system keychain, alongside more than 50 remote MCP servers and new models in the pickers, including GLM-5.3-Flash and MiniMax M3. Its updaters also now pause gateways instead of killing them. A patch release, 0.21.1, followed on September 7 with 5,139 non-merge commits and 632 merged pull requests, and Nous Research said full notes for that window will ship with version 0.22.0. On GitHub the project showed 244k stars on September 10.
Running Hermes Agent on a Mac: what StrideNote found
StrideNote runs Hermes Agent on its test stack, set up as described in our guide to an always-on assistant with Hermes Agent and LM Studio. On September 10 the installed build, version 0.21.1, matched the upstream code. The machine runs the command-line tool and the messaging gateway but not the desktop app, so Bot Mode is not available there.
The command list includes a migration tool that imports settings, memories, skills and API keys from OpenClaw, a competing open-source agent. Hermes keeps API keys on that machine in a plain file readable only by the user account, since keychain encryption is opt-in.
What the Hermes Agent Bot Mode documentation leaves out
The Bot Mode documentation does not describe support for Telegram or other messaging gateways, and it lists no cost controls or permission model for bot rooms, where each bot may call a paid model. The curated notes for everything after 0.21.0 are not yet published.
Sources: Hermes Agent 0.21.0 release notes; Bot Mode documentation; Hermes Agent 0.20.6 release notes; Hermes Agent 0.21.1 release notes.