Hermes Agent: Open-Source AI Agent with Memory

Nous Research released Hermes Agent 0.19.0, named Quicksilver, on July 20, 2026. The release focuses on responsiveness, approval handling, secret storage, subagent visibility, and delivery reliability rather than introducing one headline agent capability.

What changed

Nous reports that first-turn initialization fell from about 4.3 seconds to about 0.9 seconds, an approximately 80 percent reduction across the CLI, gateway, TUI, desktop, and cron surfaces. Reasoning output now streams by default, while the desktop and terminal interfaces received separate rendering and session-switching improvements. These figures come from the project's own release notes and should be treated as vendor-reported engineering results, not an independent WhatAI benchmark.

The most consequential governance change is that smart approvals are now the default for flagged commands. An independent model reviewer can assess a dangerous command instead of asking the user to approve every match manually. Each verdict applies only to the exact command being considered. Hermes also added user-defined deny rules that remain effective in yolo mode, plus a way to explain why a command was refused so the agent can try a safer path.

Secret handling improved through new Bitwarden and 1Password sources. API keys can be resolved from supported vault references rather than remaining in a plaintext environment file. The release also adds live transcript files for delegated subagents, making background activity easier to inspect while it runs. Nous says completed delegated work and outgoing results now use more durable ledgers so process restarts are less likely to lose a finished response.

Who benefits

Frequent desktop and terminal users should feel the startup and rendering changes most directly. Developers using subagents gain a clearer record of tool calls and responses. Teams with several providers or connected services gain a safer route for managing secrets. Users who previously disabled approvals because prompts became repetitive may find the smart-review system more practical.

What the release does not solve

A model reviewer is still a model, not a proof that a command is safe. It may misunderstand intent, miss an indirect consequence, or approve something that becomes dangerous in the current environment. Password-manager integration reduces plaintext exposure but does not make broad credentials harmless. Live transcripts improve visibility only when somebody reviews them.

The release also does not change the central cost model. Hermes remains free software, while inference, hosted tools, cloud compute, and external services can still consume money. Faster execution can increase consumption when users schedule more jobs or run several subagents at once.

WhatAI's take

Version 0.19.0 makes Hermes easier to operate, but its buying case remains ownership and extensibility. The useful question is not whether it can act without supervision. It is whether you can give it a narrow job, visible evidence, limited tools, scoped credentials, a reversible output, and an approval rule that matches the risk.

For new users, the safest sequence is to begin with read-only local work, inspect memories and generated skills, add file writes with checkpoints, then connect one external service at a time. Scheduled writes and unattended messaging should come only after the workflow behaves predictably in manual sessions.

Source: Nous Research's official Hermes Agent 0.19.0 release notes, July 20, 2026.

WhatAI Decision Box

Hermes Agent is a free, open-source agent runtime from Nous Research. It combines persistent local memory, reusable skills, model choice, terminal and file tools, messaging gateways, scheduled work, subagents, MCP connections, and desktop access.

Hermes Agent's clearest differentiator is continuity around the model. Instead of treating each chat as an isolated exchange, it can retain relevant memories, search earlier sessions, create procedural skills from completed work, and reuse or improve those skills later. It is also model-flexible: users can connect Nous Portal, OpenRouter, OpenAI, Anthropic, Google, or an OpenAI-compatible local endpoint. The core software is free under the MIT licence, but inference, hosted tools, cloud servers, and external services can still create costs. Nous Portal currently offers a Free tier, Plus at $20 monthly, Super at $100 monthly, and Ultra at $200 monthly. Hermes is best understood as infrastructure for building and operating a personal agent. It requires more setup and security judgment than a simple browser chatbot, especially when terminal access, connected accounts, scheduled tasks, or unattended gateways are enabled.

Hermes Agent brings persistent memory, evolving skills, tools, model choice, messaging, and automation into one open-source system.

Join the conversation below to share your setup, explain which model and execution backend you use, compare Hermes with another agent, report limitations, or show a workflow that became more useful over time. All constructive feedback is welcome.

Use Cases

Pricing varies by plan and region — see current pricing.

Plan features change — last updated: 2026-07-29.

Hermes Agent Pros & Cons

How to Get Results with Hermes Agent: Step-by-Step Workflow

1. Choose the narrow first job

   Define one repeatable task, its approved inputs, required output, stopping condition, and the actions the agent must never take.

2. Select the deployment boundary

   Choose local desktop, terminal, Docker, remote host, or cloud execution based on where the files, credentials, and tools can safely live.

3. Configure one model provider

   Start with one provider or a compatible local endpoint, confirm context and tool support, and set a cost limit before adding routing complexity.

4. Restrict tools and credentials

   Enable only the tools required for the task, use narrowly scoped secrets, keep dangerous-command approvals active, and prefer a sandboxed backend for risky work.

5. Supply durable context

   Add concise project instructions and verified reference material, then review what Hermes stores as memory instead of assuming every retained detail is correct.

6. Run manually and inspect

   Complete several supervised runs, examine commands, file diffs, costs, tool outputs, generated skills, and failure recovery before expanding access.

7. Make outputs reversible

   Use checkpoints, drafts, branches, logs, and approval queues so an incorrect result can be inspected and undone without reconstructing the original state.

8. Automate gradually

   Add one schedule, webhook, messaging gateway, MCP server, or subagent workflow at a time, monitor it, and keep external writes behind approval until the evidence supports broader autonomy.

Hermes Agent Gotchas and Limits to Know Before You Start

• Hermes Agent is agent software, not the underlying language model. Output quality depends heavily on the selected provider and model.
• The core software is free, but inference, hosted tools, cloud servers, messaging, storage, and external APIs can create costs.
• Self-improving refers to memories and procedural skills, not autonomous training of model weights.
• A stored memory can be wrong, outdated, oversimplified, or retrieved in the wrong context.
• An automatically created or revised skill can preserve a flawed process until a user inspects it.
• Terminal, file, browser, messaging, MCP, and webhook access create meaningful prompt-injection and data-exposure risks.
• Smart approvals reduce repetitive prompts but cannot prove that a flagged command is safe in its full environment.
• Yolo mode bypasses dangerous-command approval prompts and should not be used as a convenience default.
• Scheduled jobs run in fresh sessions, so their prompts must contain the context and constraints required for that run.
• Unattended automation can repeat an error, consume credits, or affect external systems before a person notices.
• The official FAQ says local agent data is not collected by Hermes, but requests sent to a configured model provider follow that provider's data practices.
• Local models avoid hosted inference calls but require suitable hardware, context configuration, model serving, updates, and security.
• Tier 1 platform support does not mean feature parity. Some capabilities are unavailable on native Windows or Android.
• MCP servers, skills, plugins, hooks, and third-party integrations bring their own code, permissions, dependencies, and supply-chain risk.
• Messaging gateways must restrict who can contact the agent and what each identity is allowed to trigger.
• Cloud and remote terminal backends can retain files or sessions differently, so storage, shutdown, and billing behaviour need verification.
• Checkpoints help with file recovery but do not automatically reverse emails, messages, API writes, purchases, or other external actions.
• Subagents can accelerate work and multiply token, tool, infrastructure, and review costs at the same time.
• The project evolves quickly, so commands, platform support, providers, features, and defaults can change between releases.
• Pricing, included credits, rollover caps, rate limits, and model availability should be rechecked before paying or automating at scale.

Which Hermes Agent Feature Fits Your Use Case

Starter Prompts for Hermes Agent

What is Hermes Agent?

Hermes Agent is a free, open-source agent runtime from Nous Research. It can use language models through several providers, operate through desktop, terminal, messaging, and programmatic surfaces, and combine memory, skills, tools, scheduling, and delegation.

Is Hermes Agent free?

The Hermes Agent software is free under the MIT licence. A user still needs inference from a local model or an external provider. Local inference can avoid API charges, while hosted models, tools, cloud servers, messaging services, and other infrastructure may cost money.

How much does Nous Portal cost?

The current Nous Portal plans are Free at $0, Plus at $20 per month, Super at $100 per month, and Ultra at $200 per month. Paid tiers include monthly credits, access to hundreds of models, hosted tool usage, and higher rate limits. Usage beyond included credits and cloud server costs can add to the total.

Which models work with Hermes Agent?

Hermes supports Nous Portal, OpenRouter, OpenAI, Anthropic, Google, and several other configured providers. It can also connect to local or self-hosted inference through Ollama, vLLM, llama.cpp, SGLang, or another OpenAI-compatible endpoint.

What does self-improving mean in Hermes Agent?

It means the agent can preserve memories, create reusable procedural skills from experience, and revise skill instructions during later work. This is not the same as autonomously retraining or changing the weights of the language model that powers a session.

Where does Hermes store my data?

The official FAQ says conversations, memory, and skills are stored locally in the user's Hermes directory and that Hermes does not collect telemetry, usage data, or analytics. Model requests still go to the provider the user configures, so the provider's data terms and the content sent in each request remain relevant.

Can Hermes Agent run offline?

Yes, when connected to a compatible local model endpoint. Features that depend on hosted models, web search, browser services, messaging networks, cloud infrastructure, or external APIs will still require network access.

Can Hermes Agent edit files and run terminal commands?

Yes. Hermes includes terminal and file tools, can create checkpoints before file changes, and supports several execution backends. Commands and edits can still cause damage or expose data, so users should keep approvals, write controls, sandboxing, backups, and limited credentials in place.

Can Hermes Agent run scheduled tasks?

Yes. Hermes supports one-time and recurring jobs through natural language or cron expressions. Jobs can attach skills and deliver results to files or configured messaging platforms. It also supports script-only jobs that do not invoke an LLM.

Does Hermes Agent support MCP?

Yes. Hermes can connect to external MCP servers over supported transports, including stdio and SSE. The available tools, credentials, security boundaries, and reliability then depend on both Hermes configuration and the MCP server being used.

Which operating systems are supported?

The current Tier 1 list covers Apple Silicon macOS, Windows 10 and 11, Linux and WSL2, and Docker containers. Android through Termux and Nix are maintained on a best-effort Tier 2 basis. Some features are unavailable on native Windows or Android, and Intel macOS is not supported.

Is Hermes Agent safe to run unattended?

Hermes includes approval modes, write controls, sandbox options, credential filtering, context scanning, and session isolation. These reduce risk but do not remove it. Unattended jobs should use narrowly scoped tools and credentials, deny dangerous commands by default, and deliver reversible drafts before making external changes.

How is Hermes Agent different from Claude Code or OpenAI Codex?

Claude Code and Codex focus strongly on software work around their respective model ecosystems. Hermes is an open-source, model-flexible runtime that also emphasizes cross-session memory, evolving skills, messaging gateways, scheduled automation, local-model support, and multiple execution backends. The better choice depends on whether model quality and coding workflow or agent ownership and extensibility matter more.

Related Agents & Automation Tools

Explore the Network