One word conclusion
Hermes Agent is a "long-running, cross-platform, memorized, and skilled individual/team AI Agent operation layer" Research by Nous ". It is not an Agent SDK that only embeds code for developers, but is more like a AI assistant runtime that can be deployed in this machine, VPS, container, SSH remote machine, Modal/Daytona cloud sandbox: users can talk to it at CLI, desktop, Telegram, Discord, Slack, WhatsApp, Signal, Email, enterprise IM, API Server, IDE and other portals, it calls models, tools, MCPs, terminals, files, browsers, scheduled tasks, and sub-agents behind the scenes to complete the work.
For pre-sales, what deserves most attention is not "another chat robot", but three things: first, cross-channel unified Agent portal; Second, Agent forms a closed loop of sustainable learning through Memory, Session Search and Skills. Third, it has a relatively high degree of engineering for tools, MCP, container isolation, approval, security boundary, background tasks and timing tasks, it is suitable to take the landing form of "enterprise-level Agent from Demo to resident operation.
What the project can do
1. AI Agent Across Terminals and Messaging Platforms
Hermes has two main types of portals: one is the local/remote work portal, including CLI, TUI, desktop, API Server, IDE/ACP adapter; The other is the message gateway, including Telegram, Discord, Slack, WhatsApp, Signal, Email, Home Assistant, Mattermost, Matrix, DingTalk, Flying Book/Flying Book International Edition, Enterprise WeChat, WeChat, QQ, Teams, LINE, ntfy, Raft, IRC, etc.
This means that users do not have to open a Web application to use the Agent, but can initiate tasks from their existing work portal. For example:
-send a "help me summarize the competition news at 8 o'clock every morning" in the Telegram ".
-In Slack, let it check the status of a service, organize logs and send daily reports.
-Let it change code, run tests, and read local files in CLI.
-Use Hermes as the editor agent in the IDE.
-Connect it to the enterprise automation process in the API Server or Webhook.
In pre-sales expression, it can be understood as "one Agent, multi-terminal entrance and one background gateway to undertake tasks in a unified way".
2. Multi-model and multi-service provider routing
Hermes supports Nous Portal, OpenRouter, OpenAI, custom OpenAI-compatible endpoint, and a variety of model providers mentioned in the documentation. Users can switch models through 'hermes model' or '/model provider:model' instead of rewriting the call logic in the business code.
The value to the customer is:
-Avoid being bound by a single model vendor.
-You can cut models by task, such as encoding, long text, and low-cost background tasks.
-You can connect the existing model gateway or private deployment interface of the enterprise.
-Nous Portal also makes model, Web Search, image generation, TTS, cloud browser and other tool capabilities into a unified subscription portal to reduce API Key configuration costs.
3. Tool System: From Dialogue to Real Execution
Official documents show that Hermes built-in tools cover Web search, web page extraction, terminal execution, file reading and writing, patch, browser automation, visual analysis, image generation, TTS, TODO, memory, session search, timing tasks, sub-Agent delegation, Home Assistant, MCP external tools, etc. The architecture document mentions that the inside is a centralized Tool Registry, with about 70 tools and about 28 toolsets.
Typical capabilities include:
-'terminal': Execute commands and support background processes.
-'read_file/'patch': Read and modify files.
-'web_search'/'web_extract': Search and extract web content.
-browser_navigate/browser_snapshot/browser_vision: Browser automation.
-'delegate_task': Delegates sub-Agents to process tasks in parallel.
-'cronjob': create, run, pause, resume scheduled tasks.
-'memory'/'memory': session_search long-term preferences and search history sessions.
-MCP tools: connect GitHub, database, internal API, file system, third-party SaaS and other external tool servers.
Pre-sales point: Hermes is not "only reply", but has a task execution link. Its key value is to string the LLM, tools, execution environment, and message entry into a long-running Agent Runtime.
4. Memory and skill closed loop
Hermes's README positions itself as a "self-improving AI agent". Its learning loop is mainly composed of three parts:
| Module | Role | Pre-Sales Explanation |
|---|---|---|
| Persistent Memory | Save user preferences, environmental facts, project conventions, etc. | Let the Agent understand user and team habits step by step |
| Session Search | Retrieves historical sessions based on SQLite FTS5 | allows the Agent to recall specific content discussed in the past |
| Skills System | Precipitate repetitive processes into reusable skills | Turn one-time experience into SOP that can be called next time |
Memory documents show that by default, Hermes uses 'MEMORY.md' to save Agent's own work notes and 'USER.md' to save user portraits, and there is a character limit to avoid unlimited memory expansion. Session Search saves all CLI and messaging sessions and can search history on demand. Skills are on-demand knowledge documents that are compatible with the 'agentskills.io' open standard and support '/learn' to generate new skills from documents, directories, URLs or just completed workflows.
This is an easy value story for corporate customers: traditional Chatbot have to reinterpret the context every time, while Hermes wants to gradually precipitate "customer preferences, organizational processes, project constraints, and common fixes" to form reusable knowledge assets.
5. MCP Integration and Scalable Tool Ecosystem
Hermes can be used as an MCP client to connect to stdio or HTTP MCP Server. Hermes can also be used as an MCP server in the form of 'hermes mcp serve' to expose its message session capability to MCP clients such as Claude Code, Cursor, and Codex.
MCP is important for pre-sales because it saves Hermes from having to write a separate native tool for each enterprise system. GitHub, Linear, Stripe, database, internal API, file system, knowledge base, etc. existing in the enterprise can be automatically discovered and registered as tools by Hermes as long as they are exposed through MCP. Hermes also supports per-server tool filtering, such as exposing only list_issues, create_issue, and blocking high-risk actions such as deletion and refund.
Suitable for customer speech:
If the customer is already building the MCP tool layer, Hermes can be used as a multi-channel Agent portal to bring MCP tools to CLI, IM, scheduled tasks, desktop and automated processes.
6. Background Tasks, Sub-Agents and Timing Automation
Hermes supports background tasks and child agents. The latest Release 'v0.17.0 'emphasizes delegate_task(background = true): the main Agent can assign long tasks to the back-office sub-Agent and continue to process the current dialogue. after the task is completed, the result will return to the session.
Message gateway also supports '/background
This kind of ability is suitable for demonstrating the transition from question-and-answer assistant to automated staff.
Schematic of official architecture
The architecture diagram in the project document is a text diagram with the following core structure:
Entry Points
CLI / Gateway / ACP / Batch Runner / API Server / Python Library
|
v
AIAgent (run_agent.py)
Prompt Builder
Provider Resolution
Tool Dispatch
Compression & Caching
Tool Registry
|
+--> Session Storage: SQLite + FTS5
|
+--> Tool Backends:
Terminal: local / Docker / SSH / Singularity / Modal / Daytona
Browser
Web
MCP
File / Vision / MediaAs can be seen from this figure, Hermes's design is "separation of platform portal and Agent Core": CLI, message gateway, Cron, ACP, API Server and other portals will eventually enter the sone' AIAgent' core, and then the core will do Prompt construction, model routing, tool scheduling, session persistence and execution environment selection. This architecture is more conducive to unified configuration and unified governance than "one bot per platform.
Applicable Scenarios
Scenario 1: Resident Coding and Operation and Maintenance Agent of R & D Team
Suitable for customers: R & D team, platform engineering team, DevOps/SRE team, AI native startup team.
Can do:
-Let the Agent read code, change code, and run tests in CLI/desktop.
-Trigger inspection, deployment, log analysis, and service health reports in Slack/Feishu/Enterprise WeChat.
-Use Cron to run daily build reports, dependency risk scans, and Issue summaries.
-GitHub, Linear, Jira, internal monitoring system via MCP.
Pre-sales value: Precipitate fragmented operations in R & D and O & M into natural language schedulable processes to reduce "human flesh check status, human flesh summary, and human flesh repeat scripts".
Scenario 2: Enterprise internal multi-channel AI assistant gateway
Suitable for customers: Enterprises that already have multiple IM channels, cross-regional teams, and want to put AI capabilities into the daily entry of employees.
Can do:
-An Agent simultaneously accesses Slack, Teams, Flying Books, Enterprise WeChat, Email, etc.
-Set different Toolset and permissions for different platforms.
-Control who can access via allowlist, DM pairing.
-Deliver automated results back to the user's channel.
Pre-sales value: do not force employees to switch to a new platform, AI the ability to directly enter the existing collaborative environment.
Scenario 3: "Long-term personal AI assistants" for individuals or small teams"
Suitable for customers: founder, consultant, researcher, pre-sales, technical leader.
Can do:
-Remember user preferences and ways of working.
-Search past conversations to retrieve previously discussed projects, clients, and solutions.
-Learn common processes into Skills.
-Resident operation on VPS, remote call from mobile phone messaging software.
Pre-sales value: Emphasize that "the AI assistant is no longer just a chat window, but a personal work layer that follows the user's work for a long time".
Scenario 4: Agent Tool Ecology and MCP Landing Test Field
Suitable for customers: Teams building internal Agent platform, MCP Server, unified tool gateway.
Can do:
-Verify that MCP tool exposure is reasonable through Hermes.
-Test the stability of different models to tool calls.
-Compare local, Docker, SSH, Modal, Daytona, and other execution backends.
-Prepare data for tool invocation model training with trajectory generation and compression capabilities.
Pre-sales value: Hermes can be used as an "experience terminal" and "verification platform" for the Agent tool layer to help customers see MCP/instrumented business effects faster.
Not suitable for the scene
Hermes is not suitable for all projects. The following scenarios should be cautiously recommended:
-Only need to embed a lightweight chat component in the product, Hermes may be too heavy.
-Strong regulatory environment requires clear enterprise-level RBAC, audit, approval process, and tenant isolation. Hermes is currently more single-tenant/personal Agent and requires additional packaging.
-Scenarios that do not allow the Agent to execute commands, read and write files, or call external tools will weaken the maximum value of Hermes.
-Customers want to purchase out-of-the-box commercial SaaS, not deploy and maintain an open source agent Runtime.
-Teams that do not have a review mechanism for third-party Skills, Plugins, and MCP Server should not directly open high-authority tools.
How to Use
Quick installation
Linux, macOS, WSL2, Termux:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.bashrc
hermesWindows PowerShell:
iex (irm https://hermes-agent.nousresearch.com/install.ps1)Common commands:
hermes # 启动交互式 CLI
hermes model # 选择模型提供商和模型
hermes tools # 配置启用工具
hermes setup # 完整设置向导
hermes gateway # 启动消息网关
hermes update # 更新 Hermes
hermes doctor # 诊断环境问题Start Message Gateway
hermes gateway setup
hermes gateway startCommon commands in the messaging platform:
/new 或 /reset 开启新会话
/model 切换模型
/skills 查看或调用技能
/background 启动后台任务
/stop 中断当前任务
/reload-mcp 重新加载 MCP 工具
/usage 查看用量Configure MCP
Minimum stdio MCP example:
mcp_servers:
filesystem:
command: "npx"
args: ["-y", "@modelcontextprotocol/server-filesystem", "/home/user/projects"]Minimum HTTP MCP example:
mcp_servers:
company_api:
url: "https://mcp.internal.example.com/mcp"
headers:
Authorization: "Bearer ***"Tool filtering example:
mcp_servers:
github:
command: "npx"
args: ["-y", "@modelcontextprotocol/server-github"]
env:
GITHUB_PERSONAL_ACCESS_TOKEN: "***"
tools:
include: [list_issues, create_issue, update_issue]
prompts: false
resources: falseSecurity and deployment points
Hermes's security document is quite frank: it regards "operating system-level isolation" as a real security boundary. Approval gates, command scanning, output desensitization, Skills Guard, etc. are all heuristic protection and should not be regarded as strong isolation.
Customers must be reminded before sales:
-By default, the local backend will execute commands on the host and is not suitable for accessing uncontrolled input sources.
-In production or multiplayer scenarios, we recommend that you use Docker, Modal, Daytona, and SSH remote machines to isolate the backend.
-Message gateway must be configured with allowlist or DM pairing, not open to all users.
-MCP Server, Skills, Plugins all pose supply chain risks and need to review sources and exposure tools.
-For sensitive systems, a tool whitelist should be adopted, giving priority to open only read operations or low-risk operations.
-High security scenarios should consider container/sandbox packaging at the entire Agent process level, rather than just isolating terminal.
The official recommended production gateway Checklist include: setting clear allowlist, using container backends, limiting resources, properly storing keys, enabling DM pairing, reviewing command allowlist, limiting working directories, running non-root operations, monitoring logs, and updating regularly.
Pre-sales explanation framework
Customer pain points
-The AI assistant is currently stuck in the chat window and cannot enter the real workflow.
-Each channel has a separate bot, knowledge and conversation are fragmented.
-Agent Demo is easy, long-term operation, permissions, security, memory, tool management is difficult.
-There are many internal tools, but it lacks a natural language unified entrance.
-Employees have many repetitive tasks, such as checking data, sorting reports, running scripts, inspecting, summarizing and sending notices.
The value proposition of Hermes
-Multi-portal: CLI, desktop, IM, Email, API, IDE can be accessed.
-Multi-model: supports multi-service providers and custom endpoint to reduce vendor lock-in.
-Multi-tools: built-in Web, terminal, file, browser, media, scheduled task, sub-Agent, MCP.
-Learnable: Memory, Session Search, Skills let it gradually precipitate users and team processes.
-Deployable: native, VPS, Docker, SSH, Modal, Daytona, Singularity can run.
-Can be governed: dangerous command approval, gateway allowlist, MCP tool filtering, container isolation, SSRF protection, context injection scanning, etc.
Differences from General Agent Framework
| Comparison dimension | General Agent SDK | Hermes Agent |
|---|---|---|
| Main Location | Write code for developers to integrate agents | Run a resident agent for users/teams |
| Entry | Usually API or in-app component | CLI, Desktop, IM, Email, API, IDE |
| Long-term memory | Need to do it yourself | Built-in Memory and Session Search |
| Workflow precipitation | Need to design your own | Skills system and '/learn' |
| Tool Execution | Dependent Developer Access | Built-in Tools and MCP |
| Deployment form | Common server applications | Local, container, remote, cloud sandbox, gateway |
| Pre-sales selling points | Development flexibility | Resident operation and real workflow landing |
PoC recommendations
PoC Goal 1: R & D Assistant in Enterprise IM
Verification content:
-Access Slack, Flying Book or Enterprise WeChat.
-Configure allowlist to allow only test users.
-pick up a GitHub or internal Issue MCP.
-Let Hermes complete the closed loop of "check Issue, summarize PR, generate daily report and create to-do.
-Open the Docker backend and verify command execution isolation.
Success Criteria:
-Users can initiate tasks in IM and receive results.
-Agent can call at least one external tool.
-Dangerous commands trigger approval or are quarantined.
-A regular daily report can be automatically delivered to a designated channel.
PoC Goal 2: Pre-sales Research and Data Collation Assistant
Verification content:
-Search competing products and open source projects using Web Search / Web Extract.
-Fixed analytical templates with skills precipitation.
-Use Memory to record pre-sales preferences, such as output structure, customer industry, and common directories.
-Automatically generate competition monitoring weekly with Cron.
Success Criteria:
-When analyzing the same project twice, the output style and structure are more stable.
-Can search history sessions to retrieve previously done project analysis.
-Can learn a process into reusable Skill.
PoC Goal 3: MCP Tool Gateway Verification
Verification content:
-Access a read-only database MCP or internal API MCP.
-Set tool include/exclude to mask write operations.
-Call the same MCP capability in CLI and IM respectively.
-Observe tool call logs and failure handling.
Success Criteria:
-Agent can correctly select MCP tools in natural language.
-Permission filtering takes effect.
-The same capability can be reused across inbound ports.
List of major risks and issues
-Open source projects are large in size and have a wide range of functions. Deployment and maintenance require the input of technical teams.
-GitHub API currently shows a high number of Issue. It is recommended to verify the quality and maintenance rhythm of the issue before pre-sales presentation.
-Multi-platform gateways mean that the credentials and permissions are expanded, and customers need to be clear about who can access and which tools can be invoked.
-Skills and Plugins security boundaries rely on manual review and are not suitable for ungoverned installation of community extensions.
-For strong compliance customers, enterprise-level audit, RBAC, approval process, log retention, data isolation and other solutions need to be supplemented.
-If the customer only wants to be an Agent in a single business application, Hermes may not be the lightest choice.
Can be used for pre-sales customer questions
-Where do you want your employees to use the AI from? Web, IM, IDE, CLI, or multiple portals?
-Does your AI assistant need to call internal systems, execute scripts, read and write files?
-Do you have an MCP Server or internal tooling API in place?
-Is the Agent required to remember team processes, customer preferences, and project engagements?
-Is there a regular task scenario, such as daily report, weekly report, inspection, and competition monitoring?
-What untrusted input sources does the Agent contact, such as public web pages, external emails, and group chat messages?
-What are the security requirements for command execution, file access, credential management?
-Is it a personal efficiency tool, a research and development assistant, or an enterprise-level unified Agent gateway?
My pre-sales judgment
Hermes Agent is suitable to be put in the scheme of "Agent landing from chat to execution. It is not the lightest Agent SDK, but it puts many engineering problems required by a real running Agent on the table: multi-terminal entry, tool system, memory, skills, historical retrieval, background tasks, Cron, MCP, model routing, container/remote execution, security approval, gateway authorization, deployment service.
Hermes is a good reference project and PoC candidate if the customer has made it clear that they want to be "internal AI assistant portal", "R & D/O Agent", "multi-channel automation assistant" and "MCP tool ecological verification. If the customer just wants to make a chat box in the product, or needs strict commercial multi-tenant back-office management, Hermes is more suitable as an architecture reference than a direct delivery.