Tips and usage for hermes-agent.nousresearch.com
A set of 24 reusable skills for Hermes has been released. These skills are ready-made work steps that an agent can load only when needed, and they cover 8 areas including research, creative work, productivity, and agent maintenance. The research skills include deep research, profile-style research on people or groups, fact checking, source tracking, document-based semantic search, news monitoring, media analysis, and YouTube topic research. The creative skills include staged image and video generation through fal.ai with cost tracking, plus free stock photo lookup through Pexels. The productivity skills include Word and Excel mail merge, travel itinerary building, decision logs, file format conversion, and a task brief that gathers the goal, context, and limits before major work starts. Each skill is self-contained, so people can copy only the ones they need. The repository uses the MIT license and has passing CI tests on Python 3.11 through 3.13.
AI can work better for business tasks when it is split into clear roles instead of being treated as one all-purpose chatbot. The useful shift is not just asking better questions, but building a workflow where each AI role has memory and job-specific background material. AI still does not carry responsibility, guarantee outcomes, or replace human judgment. A person still needs to make the final calls. The setup uses Hermes Harness, with the different AI roles reached through Telegram. The roles also share the same file base on the user’s PC, so they are not working from separate information. The roles cover sales thinking, marketing drafts and campaign ideas, operations and admin work, legal and compliance review, quality control and finance-style checks, and data or research. A main orchestrator reviews the work and sends tasks to the right role.
Hermes Agent is an open-source personal AI agent from Nous Research, built for personal productivity and coding help. Its main strength is memory: it can bring back older code, notes, and work context when they become useful again. The memory system combines BM25 keyword search with Vector search, then uses a cross-encoder reranker to put the most relevant matches first. This makes Hermes useful for finding a code snippet or related context from weeks earlier. Business automation has different needs because company records must stay exact and reliable. Atom OS uses PostgreSQL as the source of truth for facts, while LanceDB helps speed up search, so the factual record can still be queried even if the search layer fails.
Hermes Agent can be connected to an OmniRoute local endpoint instead of being tied directly to one fixed AI provider. That setup lets requests move through more than one provider path, so one provider’s usage limit is less likely to stop the whole job. The difference matters most during longer tasks. Building a page, writing a script, or planning a workflow can require many model calls, and a single-provider setup can break when that provider reaches its limit. The main idea is to give Hermes Agent one flexible route rather than one brittle path. The available details focus on the idea and claimed benefit, but they do not give exact setup settings, proof of long-term reliability, or clear cost conditions.
Hermes agent’s memory system can add new memories without noticing older memories that are stale or conflict with the new ones. Adding a timestamp to each new entry in USER.md and MEMORY.md makes it easier to tell which memories are old. Older memories that no longer match newer ones can then be cleaned up or removed. A custom plugin that added timestamps helped reduce context use because Hermes agent updates memories often. The suggested improvement is to make timestamps a default feature, and possibly add an extra step that automatically prunes stale or conflicting memories.
An n8n automation runs every day at 8 AM and calls a Hermes agent through OpenRouter to find 10 possible B2B clients. For each company, it gathers an email address, phone number, a short description of what the company does, and a note on which business processes could be automated. The structured result goes back to n8n and is written into Google Sheets. The workflow works, but the cost is too high. Even with a cheap model, the agent uses many tokens because it searches the web and reads full pages for every company. The possible fixes are to reduce tool output, use cache, and make prompts stricter; rebuild more of the flow in plain n8n by scraping business registries, Google Maps, or directories for names and contact details; then use one small LLM call per company only to write the automation idea. Perplexity Sonar API is also being considered as a search-focused alternative, but the available details do not show its real cost or quality for this job.
pxpipe is used with Hermes Agent while the language model runs locally through llama.cpp. The target machine uses an AMD GPU. The setup points to a way to connect Hermes Agent to a local model instead of relying only on a cloud model. The available details do not include commands, required versions, configuration files, or troubleshooting steps, so the exact setup process cannot be confirmed.
Based on firsthand use, DeepSeek v4 Flash worked better than DeepSeek v4 Pro for agent-style work in Hermes. The higher-end v4 Pro first seemed like the obvious choice, but v4 Flash felt more responsive because it was faster. Hermes already provides a strong structure for carrying work forward, so the extra power of a Pro model did not seem very useful for normal autonomous tasks. The cost was also low: many scheduled jobs cost about $1.65 over one week. MiMo was also tested, but it felt less capable at working on its own than DeepSeek. Deep coding was treated as a separate case, with Codex 5.5 High used on a Mac for that kind of work.
Hermes agents can sometimes repeat the same action or keep trying the same tool call many times. That can burn through tokens quickly and use up a daily budget within a few hours. This makes fully unsupervised runs hard to trust, because a person may need to watch the agent and stop wasteful behavior. One proposed fix is a reliability layer that detects loops, stops them, repairs the problem during the run, and saves the fix in a database so later sessions can avoid the same failure. A shared knowledge database could also let many people collect fixes for common agent mistakes. The open questions are whether this is mainly a usage problem, a common Hermes behavior, or something that improves with a different model.
Hermes was running on a virtual server with Holographic, Obsidian, and a Telegram gateway connected. DeepSeek V4 Flash was used to keep costs low. A daily automated health log saved Hermes replies as markdown files in a fixed Obsidian folder. After the folder layout changed to an llm-wiki setup, Hermes was told to update the automated job so new replies would be saved in the new folder structure. Hermes repeatedly said it had changed the scripts and job instructions, but the next runs still saved files in the old folder. The old folder no longer existed, yet Hermes kept recreating it and putting files there. More than 10 repair attempts did not fix it, even after asking Hermes to delete old jobs, create new ones, check whether its claimed changes were real, and create and run a verification script.
Hermes already includes some features that people often try to add through extra skills. Checking the built-in features first can prevent duplicate installs. Unneeded skills can increase token use because their schemas are included with API calls, and they add more parts to maintain or break. One clear example is scheduled work: Hermes has a built-in cron scheduler for recurring jobs. It can run daily reports, nightly backups, weekly audits, and morning briefings without manual action. Results can be delivered to connected services such as Telegram, Discord, or Slack. A recurring task can be written in plain language, such as checking email and calendar every morning, summarizing them, and sending the result to Telegram.
A Hermes Docker setup depends on files such as config.yaml, SOUL.md, USER.md, and .env. A careless edit can break a working Hermes setup and cost hours to recover. The hermes config set command may rewrite the whole file instead of changing only one setting, so unrelated settings can disappear. Letting Hermes edit its own config can also change more than requested. Editing files as root can create permission problems, which may make Hermes ignore the intended settings and fall back to defaults. A single typo in YAML can stop the config from loading, and the failure may not be obvious. The safe routine is to back up first, make the smallest possible edit, avoid editing as root, validate the file, check the diff, and only then restart Hermes. The backup should also be checked to confirm it exists and is not an empty file.
Hermes can run into trouble when it is used as one central place for several agents at the same time. Multiple agents appear to work until about 4 or 5 are running, then the system halts. New sessions also cannot be opened reliably while other work is still running. The reported machine has an AMD 7900-class setup, 64GB memory, and a 3060 12GB graphics card, so the issue is not obviously caused by a very weak computer. Claude Desktop, Codex, Cursor, and Opencode do not show the same problem on that setup. The open question is whether Hermes has a concurrent session limit, needs a configuration change, or has a bug.
AgentTransfer is a tool for people running coding agents such as Claude Code, Hermes, and OpenClaw on different computers. It reduces the need to move files by hand. Each agent gets its own email address and folder, can register itself, and can send files through share links that expire. When a file is downloaded, SHA-256 is used to check that the file was not changed on the way. Agents can also find other agents by what they can do and work together in shared spaces. One Go binary includes an HTTP API, an incoming SMTP listener, an MCP server, webhook delivery, and a command-line client. SQLite stores the state. Every action gets an Ed25519-signed receipt linked to the record before it, so the log can be checked offline for missing or deleted entries. It uses the MIT license, and the local demo shows a full handoff from one agent to another in about 30 seconds.
Hermes is running on a VPS, while VS Code Copilot is running on a laptop. After enabling the Hermes API and adding it in VS Code as an OpenAI-compatible model, Copilot does not actually change files. The chat history still shows that it planned to edit files and tried to edit them. The likely issue is the workspace location. Hermes may be acting as if the files are inside the workspace on the VPS, but the real project files are on the laptop where Copilot is running.
Octo treats AI agents like task owners instead of simple question-answer tools. A project context is given once, a task is assigned in chat, and the result is reviewed by a person before it is approved or sent back for changes. The workflow avoids one all-purpose agent and instead uses separate agents for documents, code review, and research. Each agent keeps different context and instructions, so the right agent is chosen for the job. Octo runs on the team’s own infrastructure and can use OpenClaw, Codex, Claude Code, and Hermes as runtimes. The setup is meant to keep data inside the team’s network. The repository is open source for people who want to test it.
Hermes Agent was installed on a Windows PC with the native installer, but its internet features did not work reliably. The setup used LM Studio with Gemma E4B as the brain model. At first, it could not fetch the weather, run Google searches, or learn from YouTube, Instagram, and Reddit links. The failures showed messages such as 403 errors, browser blocking, and agent permission problems. After adding a Firecrawl API key and running Hermes doctor, Google search started working at least partly, but the accuracy was still unclear. Direct page links still did not return all the needed information, so the web-reading setup seems only partly functional.
A setup with a Mac Mini M4 with 16GB of memory and a PC with an RTX 4070 Ti with 12GB of VRAM was tested with several local models, including Hermes. The practical result is narrow: 8B to 12B parameter models can handle small jobs such as collecting and summarizing morning news, sorting a downloads folder, casual chat, simple one-shot code, and checking a small built-in knowledge base. Free commercial AI services can already do those jobs easily. The harder uses are where the local setup falls short: long context, reliable tool or skill use, and back-and-forth code debugging are described as not practical on this class of model. Buying stronger hardware is also a cost problem. A single GPU with 32GB of VRAM can cost about the same as 100 months of a 20 dollar monthly subscription, which makes the hardware upgrade hard to justify for many users.
Hermes Pulpie is a plugin that changes how Hermes Agent extracts useful text from web pages. It uses Pulpie, a 210 million parameter encoder model, to remove navigation menus, ads, sidebars, and footers from HTML and turn the remaining content into Markdown. It completes the cleanup in one model pass. The stated quality is close to Dripper, with a ROUGE-5 F1 score of 0.862 versus Dripper’s 0.864, while using about one third of Dripper’s model size. It automatically uses a GPU through CUDA or MPS when available, and falls back to CPU when not. Typical pages are said to use about 80% fewer tokens after extraction. It runs locally, so it does not need an API key or a cloud service. Setup means installing the kachook/hermes-pulpie package with pip, then setting Hermes Agent’s web.extract_backend option to pulpie.
Nous Research has announced that Hy3 can be used free for two weeks on Nous Portal. The free access appears to be connected to Tencent. To use it in hermes-agent.nousresearch.com, choose `tencent/hy3:free` from the model options. This is useful timing because the free period for Step 3.7 Flash is close to ending, so Hy3 can be tested as another no-cost option.
In firsthand use, a $20-per-month Ollama subscription with Ollama Minimax-M3: Cloud worked well with Hermes Agent for content creation. The setup was used for a tennis research project that generated a Tennis Wiki and knowledge base, with heavy token use and many biomechanics research materials. Nvidia NIM later provided free API keys, making it possible to run models for $0, but Nvidia’s Minimax M3 felt less stable than Ollama Minimax-M3: Cloud. That led to considering other Nvidia models such as Nemotron 3 Super 120b A12b and Mistral Large 3 675b. The open question is whether a larger number such as 675b means better intelligence, faster output, or better results than 120b. Other providers, including Nous Research, Mistral, and OpenRouter, also offer free large language models, so the choice is not only which model to use but which provider runs the same model most reliably.
A developer describes building a tool called NeuralMind: a fully local, persistent code memory system for AI coding agents, connected to Hermes Agent through MCP, a standard way for agents to link up with outside tools. With this setup, the agent starts every session already knowing the codebase's structure, which files belong together, and which files usually get edited next, so it no longer has to re-read context it already understood. That change reportedly made sessions about 5-10x cheaper. Before this, OpenClaw had been the daily driver for months. Its strengths were a flexible agent framework, a skill system, a multi-agent setup with planner and developer bots, and Slack and Telegram integration. But maintenance became a burden: a Python 3.10 dependency required constant workarounds, the skill hub called clawhub was partly broken with about half of installs failing, scheduled cron jobs didn't survive restarts, and the memory system relied only on SQLite with no compression, causing sessions to balloon in size. Those problems led to searching for alternatives, and Hermes Agent from Nous Research stood out because of its active development.
The TinyFish plugin lets Hermes Agent use TinyFish for web search and page extraction. TinyFish Search and Fetch are currently free for agent builders, so it can be cheaper than common Hermes choices such as Firecrawl and Tavily, which charge by credits or API usage. After setup, Hermes sends both search and extraction work to TinyFish. The setup flow is to install and enable the plugin, run the TinyFish setup command, and then run a live doctor check to confirm it works. It connects through TinyFish’s hosted MCP server with OAuth when possible, and it can use TINYFISH_API_KEY with REST as a fallback. The plugin is packaged as a normal Hermes plugin instead of a fork or core change, so it is more likely to keep working across Hermes updates. It is meant to work with both the CLI agent and gateway use.
Hermes Agent has a reference sheet that gathers its official slash commands in one place. The listed date appears to be July 4, 2026. Its practical purpose is quick lookup for commands that can be typed directly into the chat box. The available information does not show the command names, what each command does, or usage examples.
When Hermes Agent runs a personal knowledge system for weeks, the number of moving parts can grow quickly. This setup included a Telegram bot, Resend for email, a kanban board, four scheduled jobs, and a knowledge pipeline that sent pages into Mnemo. The main need was not a full analytics dashboard, but one small screen that answered whether everything was working right now. The useful checks were whether the Telegram bot could be reached, whether the Resend key still worked, whether the last scheduled job succeeded, and whether the model had hit a 429 error in the past hour. The target design was a phone home-screen widget with no install, a 60-second refresh, and color-coded status. The backend used FastAPI and uvicorn on a local address, kept a 45-second memory cache, ran `hermes cron list` and `hermes kanban stats`, and read the Hermes log file at `/root/.hermes/logs/agent.log`.
The Librarian is a personal knowledge store that can work across several agent tools, including Hermes, Claude Code, Pi, OpenCode, and Codex. It can run on a local network or on a private server inside a Tailnet. Its source material stays as Markdown files, so the notes can still be read and edited outside The Librarian if the tool is abandoned later. Search uses a temporary in-memory index and combines BM25, exact phrase search, and vector search. It can store short memories, longer reference material, and handoffs between agent sessions. A configurable curator LLM can automatically choose useful things from conversations to save, while an off-the-record mode prevents selected conversations from becoming memories. Users bring their own LLM key from any OpenAI-compatible provider, and local inference is available for stronger privacy. Browser plugins for saving long web content into the reference library are still under review, and version 1.0 is presented as a stable base for testing rather than a finished final product.
On Windows, removing Hermes Agent with the normal uninstaller may leave traces behind. The uninstall screen can say the job is finished while folders in %LOCALAPPDATA%, setup cache files, old environment variables, and user PATH entries remain. These leftovers can make Windows act as if the removed program still exists. That can cause trouble when reinstalling a clean version of Hermes Agent or switching to another agent. Direct testing found that after using the standard installer and built-in uninstaller, a settings folder and setup cache were still present. The hermes command also kept working in the terminal because PATH had not been cleaned. The fix used a single PowerShell script to clean the leftover folders and path settings together.
Engram Lite is a local memory layer for Hermes agents. It lets an agent save and search memories on the user's own machine. Memory lookup does not call an LLM, so it aims to avoid hidden prompts and extra tokens just to retrieve past information. In a LoCoMo benchmark judged by Claude Haiku, it is presented as performing better than many existing memory tools that mostly depend on LLM calls. The code is available on GitHub at engrammemory-labs/engram-lite.
A Linux Hermes setup using Qwen 3.6 as a local LLM through LM Studio showed a channel-specific problem. The context window was set above 64k, and the same tools, including the terminal, were enabled for both the CLI and Telegram. In a new dashboard or CLI chat, asking Hermes to run the `date` command correctly used the terminal tool and printed the real output. Through Telegram or a cron task, the same request did not run the tool. The reply sometimes showed code for calling the tool without actually running it, or hallucinated a wrong date. The active context was under 15k, so the problem did not look like a context size issue, and the same prompt worked correctly with other agents.
Firsthand use of the Minimax plan on Hermes Agent reached about 1.82 billion tokens in 15 days on a 20 dollar subscription, without hitting a major cap. The only reported limit was a 5-hour window limit, and it reset after about 4 minutes of waiting. Cost checks against DeepSeek, OpenRouter, and other model options did not come close to the same price for that amount of use. The Mimo plan reached its limit within 3 days under similar heavy use. The model used was M3, not M2.7, and the experience was that it worked quickly, gave strong answers, and stayed dependable. The practical question is whether any other option can offer nearly 2 billion tokens for about 20 dollars with M3-level quality.