Tips and usage for hermes-agent.nousresearch.com
A software engineering student uses OpenClaw and Hermes Agent almost every day for coding, automation, and internship work. Switching between the two tools often has made it hard to tell what each one does better. The main decision questions are which tool is worth using most, whether they are built for different jobs, which one holds up better in real use, and whether one has extra features the other lacks. The different setup styles may be reducing efficiency because moving back and forth adds friction. No concrete feature comparison or final recommendation is included yet; the value is the clear checklist for deciding whether Hermes Agent should become the main tool.
A research coding setup that has used Claude for two years now needs a locally hosted model because of usage limits and confidentiality. The work includes serious software development and algorithm design, not just simple code snippets. The model options under consideration include GLM-5.2, Kimi K-2.6, and Mimo. The tool question is which agent-style coding harness works well, with opencode, PI, hermes, and Claude Code listed as possible choices. The available hardware is very strong: four Nvidia Pro 6000 Blackwell server cards and eight H100 cards. The goal is for a team to use an internal coding assistant without sending private work to GPT or Claude.
After two months of using Hermes, the practical use cases covered many daily and online tasks. Hermes was used to put an agent-related website on GitHub, manage a YouTube channel, create short videos every day, and make one longer video each week. It also handled Google Drive and email management. DeepSeek v4 was the main model choice because it was cheaper, while OpenAI was used for illustrations or tasks that DeepSeek could not handle well. A small LLM was installed on a home VM with Hermes for private data. Some cron jobs still had problems, and the Kanban setup sometimes stopped without a clear reason. What started as a fun experiment became less satisfying once there were fewer new ideas for automation and more messages about things breaking.
A workflow using Hermes agent with Gemini 3.5 Flash to partly automate a process worked well for about a week. There was some drift in the results, but it was manageable. Then the workflow became unstable: fixing one error revealed more errors, and steps that had worked repeatedly later failed in new ways. The skill.md files were edited several times because the problem first looked like an instruction-file issue. Later, the pattern looked more like a Gemini consistency or quality problem than a Hermes skill problem. The same drop in usefulness appeared when using the Gemini chat window to troubleshoot. Even a small spreadsheet task with 25 rows and 2 columns did not seem reliably handled, based on the available excerpt.
A Hermes-agent setup for on-chain arbitrage hit an immediate limit around private key handling. The goal was to let the agent use a wallet with POL, find trading gaps across DEXs, and execute trades automatically. Hermes-agent refused to accept the private key. That matters because giving an agent a private key would give it broad control over the wallet’s funds. A safer on-chain trading agent needs a separate wallet control layer, limited permissions, and guardrails instead of placing the private key inside the agent chat or memory.
Local models can be fun and useful for learning, but they can take a lot of time before they produce real work. Image-to-video, text-to-image, coding, text-to-speech, and LoRA training can all turn into long setup work around the right workflow, quantization, settings, and hardware. They can help with simple code debugging or with making a rough first version of something new. For many other tasks, free web models or cloud models may be faster and better. hermes-agent can also get stuck when paired with local models, depending on which model is used. A machine with two GPUs, one with 16GB VRAM and one with 8GB VRAM, may still struggle. Qwen 27B looks usable, but it may need Q8 and a 256k context to be practical. For medium-size tasks, a paid subscription or free web tool limits may be more efficient than spending time tuning local models.
Running many Hermes Agent workers at the same time can make it hard to see who is doing what. Hermes Kanban is a small task board built to solve that problem. It uses SQLite and Flask, has no outside service requirements, and installs with one pip install command. It reads Hermes profiles automatically, so agents appear on the board without manual setup. When a new agent profile is added, the board adds it too. Each agent gets its name and description from SOUL.md, then shows up with its own column and color. Tasks can be created through a REST API, moved in the web page by dragging, and updated live. This setup gives a team of agents, such as a project manager, developers, marketing, data analyst, and designer, one shared place to track work and reduce repeated or missed tasks.
Hermes agent can be installed on an Android phone through Termux and connected to Telegram, so agents are available from a pocket device. The setup uses one script instead of a long manual install. The script installs what Hermes agent needs, including Firefox to reduce browser-related problems. After setup, multiple agents can be grouped into a team and used to help with business tasks. Codex and Antigravity can also be installed inside the mobile setup, letting agents help write code directly from the phone. The same setup was shared with both Hermes agent users and Termux users, which makes it relevant to people who want a portable agent workspace rather than only a desktop setup.
Hermes is now available in agent-shell. The linked page is about the agent-shell 0.55 update. The available information only confirms the new availability and the update link. It does not give setup steps, login details, pricing, supported features, or limits.
Cursor Composer 2.5 Fast can struggle on a serious app with many files by repeating the same wrong fix even after corrections. Large code changes can break the app’s design or touch the wrong files. Long sessions can also cause the tool to lose earlier context and invent missing details. Claude Max costs $100 per month, which is too high for a $30 to $50 monthly budget. DeepSeek V4 Pro with Hermes Agent or Aider is being considered as a cheaper and possibly more reliable setup. The real comparison is whether these tools stay consistent on large, messy codebases, handle big refactoring better than Cursor, justify their price, and support guardrails that stop silent damage to the app’s structure.
A desktop PC built in October 2020 is struggling during the setup needed to try Hermes Agent inside a virtual machine. Ubuntu 26.04 LTS and Ubuntu 24.04 LTS were tested in VirtualBox 7.2.2, and the Ubuntu 24.04 LTS installation has been running for several hours without finishing. VirtualBox 7.2.8 also appeared to hang partway through Ubuntu installation. Kali Linux did load, but its update and upgrade process took an extremely long time. VMware Workstation was considered as another option, but it was not seen as freely available to download. The PC uses an AMD Ryzen 7 3700X, a Gigabyte Aorus X570 Pro WiFi motherboard, and Corsair Vengeance LPX DDR4 3200MHz memory. The main concern is whether the computer is too old for a Hermes Agent setup because even the Linux installation step is slow.
Browser automation agents can use up a lot of context just by reading pages. A Playwright-based MCP snapshot of one Hacker News page can take about 14,700 tokens, and the full page tree is sent again after every click. agent-browser is a Go binary built to reduce that waste. It uses chromedp under the hood and connects through MCP over standard input and output. Instead of long accessibility-tree dumps, it gives the agent compact reference lines, then returns only what changed after an action. The claimed size is about 1,200 tokens for Hacker News and about 1,250 tokens for a GitHub repository page. Commands can be written by intent, such as `act "Sign in"` or filling a field by name, and unclear matches return ranked choices instead of guessing. Each action returns a verdict, such as navigation, dialog opened, status, visible change, no visible effect, or CHALLENGE, plus the XHRs that ran.
A Hermes setup is being used to build an order page on an existing WordPress site for a service that only works in certain locations. Before checkout, the page must check whether an entered address is valid for service. The address search needs to look through a preloaded list of valid postal addresses, support partial search, show autocomplete suggestions, tolerate typos, and understand short forms such as “St” for “Street” and “Rd” for “Road.” OpenAI Codex built the first address search prototype successfully. During the move into the wider order flow, OpenAI rate limits were reached, so the Hermes agent was switched to the paid DeepSeek V4 Pro API. Both models received the same Markdown context files. The practical problem is that DeepSeek struggled with an address autocomplete integration that OpenAI handled more easily, leaving the cause unclear: either a model limitation or a prompting problem.
Hermes agent is running in a Docker container on a desktop, and the goal is to let it inspect and manage several Linux home servers. The main concern is avoiding full SSH shell access, because that would give the agent broad power to run commands on the servers. One option is to run an MCP server on each Linux server, exposing only the read and write actions that feel acceptable. Another option is to run a separate Hermes agent on each server, with limited permissions or local MCP servers, then let the desktop agent talk to those remote agents through a shared chat channel, MCP, or a similar link. The desired workflow is asking the local agent to investigate a memory usage alert on a specific server, find the likely cause, and suggest ways to fix it. Other useful abilities would include updating Docker images and configuration files on the servers.
semantic-memory is a local-first knowledge base for AI agents and retrieval-based answer systems that need lasting memory. It keeps data on the user’s machine instead of sending it to the cloud, and stores facts, documents, and document chunks in SQLite. Search combines BM25, vector search, and reciprocal rank fusion, so it can find both exact word matches and meaning-based matches. Stored items can also have typed graph edges between them, which lets the system keep relationships between pieces of information. An MCP server is included with 18 tools, and it works with Hermes, Claude Desktop, and Cursor. The default embedding option is Candle, which runs locally on the CPU in Rust and does not need a separate service, API key, or cloud account. People who already run Ollama can use Ollama for embeddings, and a Mock option exists for tests.
A solo software business can run several AI tools with clear roles. Cursor is used as the main coding workspace, with less code written from scratch and more time spent directing, reviewing, and fixing. Its multi-file context makes it useful for a real codebase, not only a small side project. Claude Code is used alongside Cursor for terminal work, complex refactors, and checking architecture before changes are made. Hermes Agent is an open-source agent from Nous Research that can be self-hosted on a personal machine without a subscription. When it solves a problem, it writes a skill doc so it does not have to work out the same task from zero next time. Its memory continues between sessions, and it can be connected to Slack to handle background work that would otherwise be done manually.
Cowork 3P can fail with a “no usable models” error when trying to use a Minimax model. The same API key may still work in Hermes Agent Desktop, so the key itself is not always the problem. Cowork 3P may be failing to discover the available models automatically. Adding the needed model entries under Models can let inference testing continue without that discovery error.
An AI agent operating setup works best when daily work, agents, memory, and automation connect in one place. Building a large system first can create confusion before anything actually saves time. A better first step is to choose one weekly task that keeps repeating and takes too long, then automate that small workflow. Claude can sit at the center because it handles instructions, files, and context well. Claude CLI can make the setup feel more like a command center than another chat window. Codex can help when the workflow needs code, scripts, fixes, or technical changes. Hermes Astros fits best when the setup needs to watch topics and track keyword movement.
This is a real hardware setup for running Qwen 3.6 35B A3B Q8_K_XL with hermes agent. The goal is to use a model believed to fit well in a 96-128GB VRAM setup. The available hardware is split across two machines. One machine has two NVIDIA RTX 3090 cards, an ASRock Rack ROMED8-2T motherboard, an EPYC 7262 CPU, and 64GB of system memory. The other has a 9900X CPU, 32GB of system memory, a Radeon 7900 XTX, and a Radeon 7900 XT. The main practical idea is that mixed NVIDIA and AMD cards may need llama.cpp with the Vulkan backend to work together. There is no confirmed success result yet; this is a setup being evaluated.
When Hermes Kanban breaks or gets out of sync, a non-technical user may end up creating extra helpers to add new tasks or check task status. If even repeated work needs this kind of workaround, the Kanban board can start to feel less like a work board and more like scheduled jobs and simple rules hidden behind a Kanban screen. The real pain point is not only Kanban itself, but the lack of a clear path for handling repeated work in a stable way. The tone is light and joking, with a positive view of Hermes and the NousResearch community.
Hermes is being used through Matrix mobile clients with Deepseek, and the main issue is how to split purpose-built agents. Separate work areas are wanted for coding, research, orchestration, and other task types, so each one can keep its own context and focus. An orchestrator should ideally be able to work across profiles. Matrix channels should also be limited so only certain profiles listen and reply, such as a security channel handled by a security-focused profile. Family members may get access to the same Hermes Matrix server, but private personal context must not leak into shared family use. Family access also must not accidentally create a path into the main agent, especially if that agent can reach sensitive credentials, keys, account details, or higher permissions. The unclear choice is whether a separate profile is enough or whether a separate Hermes instance is needed. Resource use also matters on a VPS with 2 virtual CPUs, 4 GB of RAM, and 11 GB of free storage.
A first home server setup is planned with an Intel i3 9th gen 9100, 8 GB DDR4 memory, and a 256 GB drive, bought with a one-year warranty for 12k. Adding another 8 GB of memory would cost 4500 more, so that upgrade is on hold. There is no static IP, so the setup needs a stable long-term way to allow public access without losing much speed. Jellyfin may have both a local address and a public address, and the question is whether switching addresses is needed to get full home-network speed. The operating system choice is between Ubuntu Server and the simpler ZimaOS. The main Hermes Agent concern is whether it will work the same on ZimaOS as on Ubuntu, or whether ZimaOS will limit some features. A fast setup video or guide is also needed.
Hermes Agent is built by Nous Research and is open-source under the MIT license. It can be self-hosted, so it can run in a user-controlled setup instead of only through a hosted service. Its persistent memory is meant to remember projects across separate work sessions, unlike simple chatbots that lose track after a conversation ends. It can also build skills automatically from repeated use and past experience. The same agent can run through Telegram, Discord, WhatsApp, and Slack, which makes it possible to keep using one assistant across different messaging apps. Setup is described as a 60-second process using one curl command, and the tool can switch among more than 200 AI models. It supports plain-English scheduling, such as “every night at 2am,” and subagents for handling work in parallel. It also emphasizes local data, no tracking, and avoiding vendor lock-in.
Hermes and OpenClaw agents were connected to Telegram and appeared to work well at first, including tool use and automated reasoning. By the next morning, the agents no longer replied on Telegram because the usage quota had been exceeded. The API bill rose to about six times the expected cost. A likely cause was a task loop or repeated reading of the same files. Reading raw logs by hand was described as too painful to be useful. A proposed fix was to place a proxy between the app and the large language model provider so prompts could be compressed, with several cache layers including an L1 exact-match cache for repeated requests.
Hands-on testing across AI agents and tools such as OpenClaw, Hermes, Antigravity, Codex, Claude Code, Claude, and Grok Build led to a simpler view of automation. The goal was to save time, but the setup became more work to maintain than the help it gave. OpenClaw used millions of tokens, needed repeated context, and sometimes gave the wrong answer with confidence, apologized, and later repeated the same mistake. Memory often felt less like a special feature and more like saving clear rules and notes into structured .md files. Hermes was more reliable and broke less often, but it still used too many tokens for relatively small tasks. That made it hard to justify financially. Most automation may only need memory, connectors such as MCP, skills or tools, basic logic, and a good LLM.
AnySearch can be installed for Hermes but still not be chosen by Hermes at first. The skill folder is placed at `C:\Users\24442\.hermes\skills\anysearch`, with `SKILL.md`, `scripts`, `.env.example`, `README.md`, and `runtime.conf.example` inside. A `.env` file also contains the AnySearch API key. Running `python .\scripts\anysearch_cli.py search "hello world" --max_results 1` in PowerShell returns a normal search result, so the skill works from the command line. Inside Hermes, asking for AnySearch first leads Hermes to say it does not have an “AnySearch” tool and to try `web_search` instead. In another attempt, an error like `skill_view not found` appears. When the instruction clearly says to use the local anysearch skill and not `web_search`, Hermes enters the local folder, runs `python scripts/anysearch_cli.py search ...`, and returns AnySearch results correctly.
A practical job-hunting setup with Hermes Agent should separate research from sending applications. The useful workflow is to search across platforms, open each job listing, check whether it is real and relevant, score it against your skills, and draft a message for review. Browser automation can help with checking job details, but using it to submit applications or send direct messages can create account risk. Platforms like Upwork, LinkedIn, Reddit, and X may treat repeated automated applications or messages as spam. A safer setup keeps human approval before anything is sent. Hermes Agent works best here as a lead finder and draft assistant, not as a fully automatic job applicant.
Hermes and OpenClaw are open-source user interfaces and clients for AI agents. They received more than 200 GitHub stars, some contributors, and feedback from the AI community. The main problem was setup. Developers could prepare a Linux server, connect with SSH, and run Docker commands, but non-developers such as founders, marketers, and writers often got stuck in the terminal before seeing the product. Hermesita is a managed VPS hosting service built to remove that setup step. It creates a private VPS in the background, then lets the user log into the web interface after buying a subscription. Users do not need SSH or Docker. Each user gets a dedicated instance instead of sharing one rate-limited SaaS service with many other users. The listed price is $16.49 per month.
Connecting GLM 5.2 to Hermes Agent can turn one keyword or topic into a chain of separate tasks. Research, writing, editing, quality checks, and video preparation can be assigned to different AI agents instead of being handled by one general agent. The goal is to move beyond publishing one article and waiting for traffic, and instead create a coordinated workflow for search content, video discovery, review, and publishing. The related examples describe using GLM 5.2 with Hermes Agent, Ollama, Kanban boards, and specialized profiles for content work, coding plans, websites, games, and automation. One setup runs GLM 5.2 locally on a Mac Studio, then lets Hermes Agent plan and carry out long multi-step tasks without sending the model work to an outside server. The practical idea is to avoid using one expensive model for every step and instead split the work across GLM 5.2, other models, and supporting tools.
Open Dynamic Workflow is an open-source command-line tool for running AI agent work from JavaScript workflow files instead of relying only on prompts. The goal is to reduce cases where prompt-only agent runs drift away from the task, spend too many credits, or become hard to edit later. A workflow file can hold the steps like reusable code, so the same process is easier to repeat and maintain. The tool is described as working with Codex, Antigravity, Lobster, and Hermes agents. Workflow scripts can run on their own, or they can be plugged into CI/CD pipelines for team code reviews or release management. After installation, the included open-dynamic-workflow skill can generate workflows, and a workflow can be run with `npx u/travisliu/open-dynamic-workflow run workflow.js`.