Nous Research's Hermes Agent Unifies Every Client Behind one Gateway

Nous Research has released Hermes Agent v0.23.0 , nicknamed the Hearth Release. It provides the first curated changelog since v0.22.0, consolidating roughly 1,650 pull requests from four consecutive patch rollups. The update is available through the CLI, installers, Microsoft Store app, Docker image, and Hermes Cloud.

Hermes Agent is an open-source AI agent that runs through a command-line interface and a gateway connected to desktop, browser, editor, and messaging clients. It stores persistent memory and identity on disk, accumulates reusable skills, and keeps conversations available across clients. Version 0.23.0 focuses on session continuity, multi-user isolation, cost controls, and recovery when a conversation takes the wrong path.

One gateway carries every client

A single Hermes gateway can now serve the CLI, terminal UI, Desktop app, ACP editor integrations, web app, and supported messaging platforms at the same time. The gateway owns the session, so a developer can start work in a terminal, resume it through Telegram, and later open the same conversation in a browser using the same server-side session ID.

• Browser workspace: The gateway can serve the complete Desktop interface, including chat, group conversations, the star map, and session controls, through a dedicated HTTP listener.
• Concurrent access: The same live conversation can appear in Desktop, the terminal UI, an ACP-compatible editor, a resumed CLI chat, or a messaging thread.
• Deferred slot use: Opening an idle chat no longer consumes a gateway execution slot. A session takes a slot when it begins its first turn.

The browser listener remains disabled until an operator enables it. Deployments that expose the interface beyond a trusted network should protect it with appropriate authentication, TLS, and network access controls.

Rewind reaches the prompt history

The new /rewind command adds structured backtracking to the classic CLI. Pressing Esc twice from an idle, empty prompt opens the rewind flow, while numbered commands provide direct control over where the conversation resumes.

• /rewind N returns to prompt N for editing.
• /rewind fork N creates a new branch immediately before prompt N .

Forking preserves the existing history while creating an alternate path, which helps when testing prompts or recovering from an agent plan that produced poor results.

Shared rooms split by sender

Discord channels, Telegram groups, Slack channels, and similar shared spaces now create an isolated session for each sender when the platform supplies a user ID. Conversation history, tool costs, and interrupt state therefore remain attached to the individual participant. Direct messages continue to use their existing chat or DM identifiers.

Teams that want one shared context for an entire room can restore that behavior through configuration. The default favors isolation, which prevents one participant’s prompts or long-running task from changing another participant’s session.

Paid tools get hard limits

The Nous Tool Gateway places services such as Firecrawl, FLUX image generation, text-to-speech providers, and Browser Use behind one credential. Version 0.23.0 adds hard spending limits at that gateway layer, capping the amount an agent can consume while pursuing a task.

These limits apply directly to unattended workloads, including scheduled jobs and Ralph loop goals that continue generating turns automatically. Operators can bound the financial impact of repeated searches, image generation, browser actions, or speech requests before leaving those workloads running.

Delegation becomes measurable

Child-agent workflows can now validate outputs against an optional JSON Schema and include per-delegation cost in their results. JSON Schema validation lets an orchestrator reject malformed responses before passing them to another tool or agent, while cost reporting exposes the expense of each delegated task.

The default limits have also increased to 250 iterations and 10 concurrent child agents. Those settings allow larger parallel jobs through Hermes workflows and the Kanban interface, while the new validation and cost data give operators more information for controlling them.

Upgrade through the existing channels

The installation paths remain unchanged. The source repository is available under the MIT license, and existing installations can update through their current distribution channel.

After upgrading, operators should review three settings: whether to enable the browser listener, whether shared rooms should use isolated or common sessions, and which spending caps should apply to paid tools.

Operational polish drives 0.23.0

The Hermes Agent repository has passed 100,000 GitHub stars amid a rapid release schedule. The Hearth Release turns that development volume into a clearer operating model: one gateway serving multiple clients, persistent sessions across those clients, per-user routing in shared rooms, enforceable tool budgets, and editable conversation history.

Developers comparing Hermes with hosted agents such as ChatGPT Agent or Claude’s computer-use features can now evaluate a more complete set of operational controls. Version 0.23.0 addresses the practical requirements that emerge after a prototype becomes a shared or unattended service.