Tips and usage for hermes-agent.nousresearch.com
DeepSeek, Kimi, GLM, MiMo, and MiniMax can answer the first question correctly, then fail with HTTP 400 on the follow-up. The problem may be the chat client, not the model. These reasoning models send a hidden field called reasoning_content, which works like internal notes made before the visible answer. The model may need that hidden field returned on the next turn so it can validate the conversation history. OpenCode, Cursor, Claude Code, and VS Code Copilot can drop reasoning_content because it is not part of the OpenAI spec. A one-turn test can pass, while real users fail on turn two. The issue has reportedly been visible since January, and the related plugin has patched 12,551 messages.
Photon in Hermes agent sent a batch of delayed iMessage replies to old setup messages that had not received responses earlier. Fast updates are useful, but this shows a risk that new features may be released before they are fully tested. Unexpected replies can make message budget tracking harder when Codex is part of the workflow. Photon sounds like a useful idea, but it may not yet be ready for the number of people using Hermes agent. In this state, it feels risky to deploy it for client work, and there is no clear reason to assume enterprise accounts will be more reliable.
Hermes Agent’s Mixture of Agents changes the usual setup where one AI model receives a task and produces the whole answer alone. Several models work on the same problem in parallel, then an aggregator combines their separate answers into one final response. The proposed setup connects Hermes, Fusion, Sakana Fugu, Agent OS, and shared memory so the workflow can be reused instead of rebuilt each time. The main claim is that people do not need to wait for the next closed frontier model if they can combine useful models that are already available. One model may catch a technical risk, another may notice a weak assumption, and the final model can keep the strongest parts. Early hands-on results were mixed: an orchestration setup with Hermes, Deepseek V4, and Claude was abandoned after weak early results, though the problem may have been poor workflow wiring rather than the idea itself. The related material makes strong claims about beating Claude, Opus, and GPT, but the supplied excerpts do not include benchmark numbers or a clear testing method.
Hermes Agent was installed on an old MacBook and connected to OpenRouter so several models could be tested. In about two days, it used roughly 10 million tokens, but the output quality was still disappointing. Telegram was connected, user-md was configured through an interview flow, and two or three cronjobs were set up through the desktop app. The jobs were simple news-fetching and summary tasks on different topics. Smaller models handled basic chat somewhat well, but performance fell apart when cronjobs needed tool calls. Qwen3.6:35B, DS V4 Flash, and Nvidia Nemotron Super 122B gave uneven results. A larger model such as GLM5.2 produced more acceptable output, but token use became very high for such basic work. Model choices and other settings also did not appear to stay saved reliably within or across sessions.
A personal nutrition MCP server was built in one evening with Claude Code because existing apps did not allow easy export of a food diary. It was made as a remote server so it could work in the Claude mobile app, then opened to the public. After 3 months, 270 people had connected it, 63% logged a meal, and 74% came back the next day. Total usage reached 10,551 meals, 4.3 million calories, 19,441 tool calls, and 35 time zones. The most-used feature was meal logging, with about 11,000 uses. Daily summaries and goal progress checks came next. The main lesson is that simple, low-effort logging mattered more than a long feature list, because people could record food in one sentence without opening another app or searching a food database.
A managed cloud hosting idea aims to keep AI agents such as Hermes running without making the user manage servers. It is meant for people who already run OpenClaw or Hermes, or who want agents to stay online all day and run on schedules or triggers. Agents would be deployed into dedicated cloud environments, so users would not need to handle a VPS, Docker, updates, or server maintenance. The service would support one agent or a team of agents working together in the same environment. The provider would handle provisioning, isolation, setup, and resource limits, while the user watches activity, usage, and spending from one dashboard. Pricing would use included allowances plus prepaid credits that do not expire, with a hard stop to avoid surprise bills.
A Hermes Agent setup connected to Deepseek through a direct API is not delegating tasks reliably. The main agent used in Telegram chat runs on Deepsek V4 Pro. Three sub-agents were created: coding-agent for coding work, config-agent for changing Hermes settings, and research-agent for research. The sub-agents use Deepseek V4 Flash. The problem is that the main orchestration agent often handles tasks by itself instead of passing them to the right sub-agent. Memory is being used through menmosney, and the main agent has been told to remember to use the sub-agents, but the delegation still happens inconsistently.
Hermes Agent was used as the coordinator for turning a large Inner Circle Trader YouTube back catalog into a searchable personal library. The target set was 679 videos across more than 30 playlists, covering mentorship material from 2016 through 2026. The goal was not just to save videos in a playlist, but to create timestamped transcripts so any concept could be searched across the whole collection. The setup used a Hermes Agent trading profile named Robin, faster-whisper tiny model, YouTube Transcript API, yt-dlp, a local markdown wiki, and cron jobs for overnight batch runs. So far, the library contains 335 transcript files, covering 428 hours of content and 368 of the 679 channel videos. Major playlists were included, such as the 2025/2026 lecture series and the 2022, 2023, and 2024 mentorship materials. The main problem was rate limits. YouTube Transcript API could process about 50 videos in 2 minutes, then become blocked for hours, while yt-dlp hit its own limit after 10 to 15 downloads. Each tool had a different cooldown period, which made it hard to know which part of the pipeline could run again at any moment.
Gemma4-31b-qat-q4 worked better than Qwen3.6-27B-MTP-q4 as a daily model for Hermes agent in this firsthand setup. It followed complex saved rules more reliably. One memory rule said to use spoken output only when the user used voice input, and to reply in the same language the user spoke. Gemma followed that rule consistently, while Qwen sometimes used spoken output after typed input and mixed languages more often. Qwen still helped with some specific hard problems when switched in for those cases. For about 90% of the custom tools in this setup, Gemma handled the work well. It also stayed strong at tool calls and coding, felt more natural in tone, and seemed better for more languages than Qwen, which felt more focused on English and Chinese.
A point-of-sale app and an analytics dashboard use one shared database structure. Each business is separated by a business ID. The goal is for an AI agent to understand references such as business A or business B and then give the right business insight for that specific business. A skill.md file with queries and descriptions is already prepared. The main choice is which model provider gives a good balance between low cost and useful intelligence. The current plan is to use Nous Research as the model provider and let one agent handle many businesses.
Using Hermes Agent for several projects or long-running tasks makes session separation important. One current workaround is to create a Telegram group with the Hermes Agent and the user, then pin that group inside a Telegram folder so all Hermes chats stay together. That setup works, but it is tedious because every new task needs another group and more manual organization. Slack has a different problem: the agent tends to reply inside the main message thread instead of posting cleanly in the channel, and multiple agents share the same context, so separate tasks can blend together. The desired setup is closer to the ChatGPT app, where pressing Command+N opens a fresh chat right away and each chat keeps its own separate context. A stronger version would also allow one-click new sessions that connect multiple Hermes computers. t3 chat has been used with Claude Code, but it may be aimed more at developer workflows than normal agentic work.
A public IP from an internet provider such as TIM may not stay the same forever. Hermes Agent was installed on a VPS, and the firewall allowed SSH access only from the home public IP. Two days later, access to the VPS failed because the home public IP had changed. The new IP also appeared to be in different cities, including Catania and Milan, even though the connection was in Lazio. The public IP kept changing every one or two days. The practical problem is not Hermes Agent itself, but a firewall rule that depends on a home IP that may be dynamic.
Nous Research added a /learn command to its open-source Hermes Agent. The command can read a folder of source material and turn it into a reusable SKILL.md file. That source material can include API docs, an unfamiliar codebase, or many config files. This makes basic skill creation much easier than writing everything by hand. Still, a skill made from public docs alone is likely to be generic. A valuable skill needs to be shaped for a specific workflow, tested against unusual cases, and optimized so it works cleanly in real use. For better results with Hermes Agent, /learn should be treated as a fast first draft, not the finished skill.
Hermes does not appear to offer fine-grained approval for each file change yet. A safer ideal flow would show a diff before any file edit or file move, then let the user approve or reject each change. For personal assistant work on a small number of high-value documents, direct edits to the main files can feel risky. One tested workaround is to let Hermes work inside Docker on a copy of the files instead of the originals. After that, rsync diff can show what changed before anything is applied to the main folder. Another possible safeguard is to back up the folders Hermes can touch with Carbon Copy Cloner, then check whether its diff view is good enough for review.
CC Switch is a desktop control panel for managing several coding agents and AI command-line tools in one place. It aims to bring Claude Code, Codex, Gemini CLI, OpenCode, OpenClaw, Hermes Agent, provider routing, skills, and settings under one roof. When many AI accounts, models, token meters, relay tools, and config files are involved, the hard part is often not whether an AI can write code. The hard part is knowing which tool, model, account, and setting is active at any moment. CC Switch may make Hermes Agent easier to use in a mixed setup where several agents and model providers are used together.
Codex already seems able to handle reminders, email-related work, documents, forms, and small personal apps. That makes it hard to see where Hermes Agent is clearly stronger, or where Codex begins to fall short. The main use cases are automated marketing campaigns, creating and reusing content, building small internal tools or MVP apps, and later testing apps that other people could use. Apps made through vibe coding may have stability or reliability limits, but the focus here is experiments and MVPs rather than production software. The useful comparison points are concrete workflows where Hermes Agent does better than Codex, plus reliability, integrations, autonomy, and ease of use.
After Hermes is installed on an Oracle free server, the agent may answer as if it does not know the current time. It can talk about cron jobs that already ran as if they are still in the future, or suggest doing something in the morning when it is already morning. When checked directly, Hermes may say it does not keep constant background time awareness and would need a script to provide the time. That matters because a personal assistant needs time to judge schedules, task order, reminders, and what should happen today.
Hermes agents can lose useful continuity over time. The built-in MEMORY.md works for hand-picked facts such as a user preference or a project tool, but it is still a flat text file. It does not track relationships between facts, it does not understand when something happened, and it can keep growing until it uses too much of the context window. Synapse is a memory system for Hermes agents that stores information as a temporal knowledge graph. It is open-source, MIT licensed, self-hosted, and designed to work as a drop-in memory provider plugin. Its main idea is that not every memory should matter forever, so some memories can fade or be forgotten on purpose.
Hermes Agent can keep running even when its results have become unreliable. The hard part is that it may not crash, show an error, or stop a task. It can still produce fluent sentences with correct grammar while the actual information is wrong. Other self-hosted services usually fail in obvious ways, such as video playback stopping, websites not loading, or storage hardware making warning signs. A language model can fail more quietly because its answers can look convincing unless someone checks the facts. In this case, Hermes Agent was running tasks around the clock and later began producing polished but false output. The main risk is a system that appears successful while silently creating bad results.
Nostics Digital used a self-hosted HermesAgent setup to prepare for a new client meeting. HermesAgent was connected to an existing espocrm system used for CRM work. HermesAgent found the espocrm API and created a skill for the connection, while a separate API user was made for HermesAgent. After contact details, company information, and the sales opportunity were entered into a HermesAgent chat, the opportunity appeared in espocrm. Company documents about strategy, purpose, branding, marketing, products, services, playbooks, and execution plans were then organized into a draft company knowledge structure using OKF. That knowledge structure was stored in a self-hosted Gitea instance. HermesAgent was also asked to check and fix content so it matched the OKF format.
MiniMax M3 Plus is advertised as a $20 monthly plan with about 1.7 billion M3 tokens per month. The unclear part is whether that monthly quota includes tokens read from cache, or whether cached tokens are discounted like they are in PAYG API pricing. In a Hermes/OpenCode-style agent workflow with a very high cache hit rate, only 2 to 3 prompts showed 43.56 million total tokens, 43.03 million peak tokens, and a 95.5% cache hit rate on the dashboard. That makes it look as if about 43 million tokens may have counted toward usage even though most of the work came from cache. The open question is whether cached tokens really reduce the 1.7 billion monthly quota, or whether the dashboard only shows processed tokens while the actual subscription quota drops by a smaller amount.
The setup is a Mac Mini M4 with 16GB of memory for running Hermes agent. The main question is whether Hermes agent requires an online subscription such as Claude or AI Codex, or whether it can work with a self-hosted model. There is a 2TB external drive available, so there is room to download a large model file. The concern is whether 16GB of memory is enough to actually run a large model well.
A one-person business can use AI tools to cover operations, planning, design, and financial tracking without hiring help for every task. Registration and tax issues still need professionals, but AI subscriptions can be worth the cost when they raise output. Choosing tools badly at first can waste money, so the useful ones are the ones that survive after real use. The tools still in use are Hermes Agent, PixVerse, Claude or Gemini, and Perplexity. Hermes Agent is used every day and can often run tasks automatically, but its permissions need to be set carefully. PixVerse is used for promo videos and posters because it puts image creation, video creation, and image-to-video work in one place. Claude and Gemini are still being compared, with Gemini improving and Claude seen as stronger for creative work. Perplexity is used mostly instead of Google Search, while Midjourney was cancelled because image restrictions became annoying.
Hermes may not play the same role as Codex or Cursor, which are mainly used for building web apps and writing code. The more practical use case is handing off repeated business tasks, such as managing social media accounts, publishing posts, replying to messages, and finding potential customers. Using only the desktop app creates a limit because the computer has to stay on for Hermes to keep working. That makes Telegram connection and running Hermes on an outside host such as Hostinger worth considering for 24/7 use. The main tradeoff is cost, so the value of the automation should be tested before paying for always-on hosting.
Hermes can do more than answer questions when a device is connected. During a phone troubleshooting session, a USB cable stayed plugged in, and Hermes used ADB commands to check battery data before answering a question about the battery chip and whether a battery might be disconnected. When a printer issue came up, the printer produced a test page. Asking Hermes to fix broken things can lead it to take real action on connected devices. It also seemed to keep improving its own skills and learning more about the user’s setup over time. The main point is that Hermes may act directly on phones, printers, and other devices, not just give advice.
Deepseek v4 flash and Pro do not directly handle images, so results can fall short when a task depends on a screen, picture, or visual layout. In Hermes, a separate vision model API can be called as a sub-agent to cover that gap. The vision model first turns the visual problem into a clear text description, then Deepseek v4 flash uses that description to produce the answer. This can create a stronger feedback loop without forcing the user to explain every visual detail by hand. Firsthand testing in Hermes and Kilo found much better results after prompting the agent to work this way. The sub-agent handoff is still not as smooth as top frontier models, so the setup may need clear instructions.
A Hermes agent used as a manager did not reliably run a hands-off development sprint. The setup already had specs, product requirements, and schemas prepared, and the manager was told not to write code. Its job was only to coordinate work, follow open-spec, and review pull requests before merging. The worker tools were claude-code and codex running in a terminal. In practice, the manager stopped moving the process forward when a worker finished a task or met a small blocker. Human supervision was still needed to restart the loop and keep work from going idle. The main weakness was session discipline: the manager did not keep checking status, deciding the next step, and pushing the workflow forward on its own.
Hermes Agent is already running correctly on a Linux machine with a local large language model. The goal is to use that same Hermes Agent from a Windows laptop on the same local network. The setup must stay fully private, so Discord, Telegram, or any outside relay service is not acceptable. The needed workflow includes pasting screenshots from the clipboard into chat and attaching files to a prompt. The current workaround uses an SSH tunnel from the Windows laptop to the Linux machine’s dashboard, but that dashboard does not support pasting or attaching files in the chat. Another possible workaround is installing the full Hermes Agent on the Windows laptop only to use Hermes Desktop and point it back to the Linux machine, but that feels indirect and not ideal.
Hermes was connected to a live geopolitical risk engine through a read-only MCP server. The important design choice was to give Hermes facts about what happened and what changed, not predictions or advice. If asked whether to buy bitcoin, it refuses instead of inventing a signal. Most of the work went into stopping Hermes from making up extra claims around real data. The same setup has run with a small local model, Claude, and DeepSeek. The model can change, while the trust rules stay attached to the data layer. For now Hermes only observes, and action-taking will be considered after a 30-day trial.
A self-hosted 24/7 personal agent is running on a Mac mini and handling both media-server work and everyday admin tasks. It manages an Unraid media stack with Plex, Sonarr, Radarr, FileBot, and live TV channels, runs Home Assistant, and downloads useful material from archive.org. It also helps with personal and work tasks such as concert alerts, backpacking trip planning, nonprofit fundraising support, long-running ancestry research, and an RSS news feed. The agent is meant to work in the background all day and only interrupt for real decisions. Conversation happens through a private Discord server, while high-priority alerts are sent through Telegram. The target workflow is one priority-ranked project tracking list for everything active or pending, plus one separate “waiting on me” list for decisions the agent needs from the person. The agent keeps working through the project list, asks questions or drafts proposals when blocked, and the person periodically reviews the waiting list to unblock more work. The setup is still early and may be rebuilt on OpenClaw or moved to Hermes.