Hermes Agent is an autonomous Agent that runs on your own machine or server. Its official gateway supports 20+ messaging platforms including WeChat, Feishu, DingTalk, WeCom, QQ, Yuanbao, and others, and each platform differs in capabilities such as voice, threads, and streaming. The Agent has persistent memory, skill self-evolution, and parallel sub-Agent capabilities, and gradually adapts to your way of working over time.
Versions and Releases The latest stable version is v0.21.6 (2026-10-08), a patch release that aggregates about 2,100 PRs since v0.21.5. The desktop, Termux, and Microsoft Store channels were not updated in sync with this release; see the Release notes for details. This is the first version to go through the new stable release pipeline, providing the Docker images nousresearch/hermes-agent:stable / :latest, and includes Dashboard authentication and security-related fixes (including items reported by Tenable Research). Capabilities not listed item by item within the window include real-time speech-to-text stt.streaming, plugin host isolation, local models / llama.cpp CUDA, Kanban unified workflow, and others. The official statement says that v0.22.0 will publish a more complete curated version update log (as of the research date, it has not yet been released). Previous versions include v0.21.5 (2026-09-24), v0.21.0 Pantheon (2026-08-31), and others; see the Releases page for details.
Usage and Installation The installation command comes from the official installation documentation. In addition to command-line installation, Hermes Desktop (Windows / macOS installers, see hermes-agent.nousresearch.com), the Android aarch64 Termux APT repository, Nix/NixOS, and others are also optional. In the configuration and start/stop examples, platform capabilities differ, unauthorized users are denied by default, and allowlists / DM pairing codes are supported. The materials include illustrative scripts explaining the gateway and automation capabilities, and do not represent real user reviews or officially promised results.
Main Capabilities For memory, it has built-in MEMORY.md / USER.md file memory (with character limits, stored in ~/.hermes/memories/) and SQLite FTS5 cross-session full-text search (session_search); optionally via Hermes memory setup integrates external vector/semantic memory plugins such as Supermemory and Mem0. Skills can self-improve during use. Built-in Cron allows scheduled jobs to be configured in natural language and delivered to configured messaging platforms; starting from v0.21.0, scheduled tasks with persistent memory and continuity are supported. Complex tasks can be executed in parallel through isolated sub-agents (delegate_task), with support for execute_code to call tools via RPC; starting from v0.21.0, mid-course steering is possible, with a default concurrency limit of 10. Hermes Desktop provides Bot Mode (multi-named Bot group chats, @mentions) and hermes peer bot direct messages.
Platforms and Deployment Supports Docker and native Windows (PowerShell installation script), Linux, and macOS; there are also 7 terminal execution backends (local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox, etc.). By default, unauthorized gateway users are denied, and security mechanisms such as command approval are provided; when using cloud models, conversations are sent to the corresponding provider. It can connect to any MCP server (desktop Connectors / MCP management page, supporting OAuth). Built-in browser automation and visual capabilities; capabilities such as search, cloud browser, image generation, and TTS can be used through Nous Portal Tool Gateway or a self-built backend. Official Windows / macOS desktop applications are provided (Simple / Advanced interfaces, built-in browser control, CLI command palette Ctrl+P, etc.); see hermes-agent.nousresearch.com for the installer entry points.
Integration and Ecosystem Compatible with the agentskills.io skill standard; the plugin directory supports SHA-pinned installation. Migration from OpenClaw is possible (hermes claw migrate). The community third-party WeChat bridge HermesClaw is an unofficial product; please verify it yourself before integrating.
Model Access In addition to Nous Portal (officially claimed 300+ models), OpenRouter, OpenAI, Anthropic, and other global providers, the official documentation integrations/providers lists multiple access methods friendly to Chinese developers. For local inference, options include Ollama, LM Studio, vLLM, llama.cpp, etc. (a context of ≥64K is recommended). By default, conversations request the API of the model provider you select; when using local models, inference can be completed on the local machine. Anthropic OAuth only applies to Claude Max and only when it includes additional usage.
Conversation Commands An excerpt from the official README shows that slash commands such as /model, /new, /retry, /compress, /usage, and /stop can also be used within conversations.
Use Cases Built-in browser automation and AI vision can be used for multi-webpage material extraction and report generation; search and image generation, etc., require configuring Tool Gateway or the corresponding provider. The materials do not specify other concrete use cases, so no further information is provided for now.
Limitations and Notes This Chinese site is an unofficial introduction site; all content should be subject to Nous Research's official channels. Data is as of 2026-10-11, compiled from official GitHub Releases and documentation. Messaging platforms differ in capabilities such as voice, files, threads, and streaming; please consult the official platform comparison table. Other features, fees, prices, platforms, effects, or usage steps not explicitly mentioned in the materials are not provided for now.
