Tips and usage for hermes-agent.nousresearch.com
The Hermes Agent buildathon on July 11 is built around making something during the day and demoing it that evening. There are no talks or panels, so the focus is on working results. Three tracks are suggested: virality, where people would want to share the result; revenue, where real signups or payments happen the same day; and AI-as-an-agency, where the tool replaces a service an agency would normally charge for. A thin wrapper around existing tools is not enough. The goal is something clear enough to make someone stop scrolling.
Hermes Agent and Paperclip can be used to move beyond one chatbot doing every task. Hermes Agent is framed as the worker that carries out jobs, while Paperclip acts as the management layer that assigns work and connects each step. The practical idea is to create specialist agents, give each one a clear role, link their outputs, and keep the whole setup focused on one main goal. This copies how a company works, with departments, responsibilities, reporting lines, deadlines, and checks that work was finished properly. The example points to using this kind of setup as an operating system for saving time, making money, and supporting repeated work. The material is more like a promotional walkthrough than a measured test, so it is best treated as a workflow idea rather than proof of strong results.
Owl Alpha is no longer available, so Hermes needs another model provider to stay useful. The need is not the strongest model, but something that works and does not hit strict rate limits too quickly. The immediate fallback is trying Google AI Studio. Unlimited use is not expected, but a service that can handle low to moderate use for about a month would be enough. Local large language models can run, but their context limit is too small for Hermes work, so they are not a practical option here.
A Hermes agent can appear ready to switch to a fallback provider when the primary provider hits a daily limit. In this case, it did not switch. Telegram first showed a working status with retry backoff, saying it was on retry 2 of 3 with 89 seconds left. Then it showed a rate-limit message telling the user to wait and try again. The agent never used the fallback provider. The fallback provider itself seemed to work, and it also appeared in the command line when running `hermes fallback`. The real issue is whether Hermes needs another setting before it will switch, or whether rate limits are handled as retries instead of fallback failures.
A local AI setup is being planned with Ollama plus Hermes Agent or OpenClaw, and the main choice is between Qwen 3.6 27B and Qwen 3.5 35B. The machine has an RX 9070 XT with 16 GB of VRAM, a Ryzen 7 7700 processor, and 32 GB of DDR5 memory. The model would mainly handle reasoning, RAG, and coordination across AI agents. Audio work would be handled by Whisper, and image analysis would be handled by a separate vision model. The practical question is whether the 35B model is worth the extra resources for long agent workflows and reasoning, or whether the 27B model or another option would be a better fit.
Installing Hermes Agent directly on the host gives it broad access to the machine, which can feel too open for regular use. Running Hermes Agent inside Docker is a way to choose exactly what it can access. The setup had trouble connecting the Docker web UI and dashboard to a Hermes install running on the host. The current problem is that the web UI or dashboard cannot read Docker logs. The container has `/var/run/docker.sock` mounted at `/var/run/docker.sock`, but Hermes Agent still says it cannot read the Docker logs. The shown setup includes a Hermes Agent container, a Hermes dashboard container, a shared Hermes home folder, an `/opt/dockge` folder mount, the Docker socket mount, and environment variables such as `HERMES_HOME`, `HERMES_UID`, and `HERMES_GID`.
Workplace security has required Hermes Agent to be stopped for now. The concern is that Hermes Agent can access the filesystem, run terminal commands, write to git, and is not on the company’s approved tool list. A security review may take months or may never happen. The practical need is to recreate as much of the Hermes Agent workflow as possible with already approved tools: Claude Code, Opencode, Warp, and Codex CLI. The main features to copy are self-modifying skills, persistent memory across sessions, and scheduled jobs that run automatically. Messaging integration is not needed for work, because the intended setup would stay local.
Metronix Core is a self-hosted memory tool for AI agents. It can pull information from Jira, Confluence, Notion, GitHub, Google Drive, Slack, and local files so an agent has more useful background to work with. It combines different search methods and a knowledge graph to find relevant material, and it can provide source citations. Its memory is split into facts, preferences, and pinned items, with separate scopes for each workspace and agent. It also checks for stale or conflicting facts, which is meant to reduce confident wrong answers. It can connect to Cursor, Claude Desktop, Claude Code, Hermes, and Open WebUI. It can run locally with Ollama, so an outside large language model is not required. The stack uses Python, PostgreSQL, Qdrant, Neo4j, and Redis, and it runs as one Docker Compose setup under the Apache-2.0 license.
Cotal is a coordination tool that lets several AI agents work in one shared space. Agents spread across tmux sessions can see the same workspace, message each other directly, and pass work between themselves. It can connect tools from different vendors, including Claude Code, opencode, Hermes, and other tools the user wires in. In the example setup, two GLM-5.2 agents handle frontend and backend work, GPT-5.5 reviews in the background, and Claude Opus leads the loop. The user gives one short instruction to Claude Opus, then Opus delegates work to the other agents without the user manually relaying messages. A top console shows the messages agents send to each other in real time. The group then builds a new console feature: a tree view that shows all agents.
For agents and automated messaging systems such as Hermes, Telegram is currently easier to use reliably than Signal. Signal requires a phone number for registration, which makes it awkward for a non-human automation account. Telegram lets people create bots that can work separately from a personal phone number. Code for sending Telegram messages is widely available, and new tools often support Telegram first, with Hermes given as an example. Signal support is often handled through unofficial tools such as signal-cli, which can be hard to set up and unreliable in practice. Even if a whole organization uses Signal, that setup can block agent alerts, automated reports, and other machine-sent messages.
Hermes and Openclaw were compared side by side for one full day of everyday work. The test used three main kinds of tasks. The first was personal assistant work, such as appointments, emails, reminders, and searching past records. One example connected a local model to a full WhatsApp chat history with a spouse, then used it to find every time a child had been sick and what medicine was given, turning that history into a table for a pediatrician visit. The second was creative developer work, where agents helped discuss ideas, make decisions, set goals, define done criteria, and compare user interface drafts made by different agents. The third was developer orchestration, where an idea moved through a repeated flow: scout, planner, planner verifier, coder, and auditor. The preferred models by role were Mimo v2.5 for scout, Opus for planner, GLM-5.2 for planner verifier, Composer 2.5 for coder, and GPT-5.5-High for auditor. Orchestration and development work ran mainly on Composer and Mimo v2.5 Pro, while design work used GLM 5.2 or Gemini 3.5 Flash.
Hermes Agent is presented as an AI agent from Nous Research that can learn from its mistakes. The practical questions are what it is useful for, how to install it, and which model to choose. The available item does not include exact install commands or specific model names. For someone trying to use hermes-agent.nousresearch.com better, the main takeaway is to decide the job first, then check setup requirements and model fit.
A monthly AI tool budget of about $200 needs careful usage planning. The current setup is split across Nous Portal, OpenCode GO, Codex 5x, and Claude Pro. Most development work happens in Codex Goals, but the allowance runs out quickly. A 5-hour limit is usually used up in about 3 hours, and the weekly limit ends with 2 to 3 days still left. Smaller tasks are being moved to Sonnet 4.6 and DeepSeek v4 Flash, but it is unclear whether that is the best setup. The main choice is between Codex 5x plus Claude 5x, which keeps model variety and leaves OAuth/API usage available for a Hermes agent through Codex, or Codex 20x, which gives higher usage limits and better reset windows while they are available.
A family with several light AI users can save money with a shared plan like Google AI Pro, which can be used by up to 5 family members. That setup becomes less suitable when the goal is a proactive agent that stays available in messengers all day and avoids being locked to one AI provider. Google Gemini subscriptions may not plug neatly into Hermes Agent, and Anthropic may not work for this sharing idea either. One possible setup is to run Hermes Agent on a $10 VPC, use a $20 ChatGPT Plus plan for the strongest model, and add Ollama Cloud or Nous as a cheaper worker model for lighter jobs. The expected total is about $50 per month for the family. Each family member would get a separate Hermes profile with a predefined set of agents, such as one stronger agent and one average helper agent. Mattermost could run on the same VPC, with separate channels for each person’s agents and channels hidden from the other family members.
Qwen3.6 35B A3B can run locally on a Ryzen 9 3900 processor, 64GB of memory, and an RTX 4060 graphics card with 8GB of video memory. The setup uses the llama.cpp server, sends many model layers to the graphics card, uses compressed cache settings, and allows a very long context window. It reaches about 22 tokens per second. The model is only acceptable for complex coding, so Composer2.5 in Cursor and Deepseek V4 Flash/Pro in VS Code are better choices for heavier programming work. Qwen is used for light coding, writing, and research, with Brave search connected inside Hermes. Tailscale makes Hermes and Open WebUI reachable from a phone and laptop. Home Assistant integration is still a next step. For people without a large hardware budget, this kind of setup can reduce the need for paid subscriptions for research, writing, and simple questions.
A carefully trimmed Hermes config can be expanded or reset when only one setting is changed from the command line. Running something like `hermes -p [argument]` may cause the whole profile config to be rewritten instead of changing only the requested option. The practical problem is that a small one-time tweak can alter the entire saved setup. This is disruptive for people who want to keep a clean custom config while using Hermes regularly.
oMLX does not smoothly unload the old model and load a new one when Hermes asks for a different model. Pi handles the same situation a little better: the first request for a different model returns an error, but a second try unloads the previous model and loads the new one. With Hermes, the switch appears to get stuck. LM Studio handles this kind of model switching more gracefully.
Fleet is a local web console for creating, configuring, watching, and operating Hermes agents that run in Docker. It is meant for people who run more than one agent on a workstation, home server, VPN, or trusted LAN. The console brings noisy daily tasks into one place, including service health, provider defaults, shared credentials, chat sessions, browser sidecars, VNC, terminal access, local web publishing, backups, restores, clones, remote nodes, and setup checks. Standard Hermes agents are the default path, and Nemo Hermes agents are supported when the nemohermes runner is available or automatic installation is enabled. Setup needs Node.js 20 or newer, npm 10 or newer, Docker with Docker Compose v2, and git if Fleet should automatically clone the Hermes source. The quick start is `npm run setup`, then `npm start`, then opening `http://127.0.0.1:5180`. Fleet keeps runtime data and secrets local by default, requires an access token when exposed on a LAN, and excludes secrets from backups unless the operator deliberately includes them.
Hermes Agent brings several slow SEO outreach tasks into one workflow. It can find relevant websites, enrich leads with contact details, prepare email drafts, and track replies. The main lesson is to build a useful prospect list before automating messages. A campaign should look for websites that already publish material related to the page being promoted, such as blogs, resource pages, or agencies in the same area. The example highlights finding 8 qualified prospects in minutes. The practical focus is not mass emailing, but turning relevant prospect research and outreach follow-up into a repeatable system.
Hermes and Claude coding agents can leave useful work behind in temporary workspaces. If the workspace is deleted after the task, small but important details may disappear too. The repository can be opened again, but uncommitted changes, scratch files, test output, and half-finished local states may be gone. Those details may not be clean enough for a Git commit, but they can still help the next agent understand what already happened. When a new agent only receives the repository and none of the working context around it, it can feel like the project is starting from zero again.
Hermes Agent can be used as the agent layer, but it still needs a model behind it to think, write, and decide what tools to use. A low-cost setup connects Hermes Agent to free cloud models through OpenRouter, while also keeping local models and existing paid subscriptions available for different tasks. The basic setup is to create an OpenRouter account, open the API key area in settings, make a fresh API key just for Hermes Agent, and paste that key into the Hermes model connection profile. Using a separate key avoids mixing Hermes with another important project and makes cost or access problems easier to trace. Related experiences point to Nemotron Ultra, Nemotron Super, and Step 3.7 Flash as free models worth trying, while Deepseek V4 Flash and Minimax M3 were reported to produce hallucinations in some Hermes workflows. For a personal computer with a 3070 graphics card, the practical question is finding a small 4 to 16B local model that can use tools correctly and write usable code. Connecting existing subscriptions such as Claude Pro through OAuth may run into token errors, so the safer approach is to keep separate profiles for free, local, and subscription-based models and switch based on the job.
Hermes Agent and LiteLLM were running on an Azure Ubuntu virtual machine, while Hermes Desktop ran on a Windows laptop through an SSH tunnel. Hermes Desktop failed to open or stay connected, and the same failure happened more than 100 times in one day. The desktop log showed connection refused, connection reset, socket hang up, a 15-second backend connection timeout, stale connection removal after a failed liveness probe, and a GPU crash loop. The likely cause is that SSH tunnel delay varied from about 200 milliseconds to 10 seconds, while Hermes Desktop used an aggressive liveness probe timeout of about 2.5 seconds. When that check failed, Hermes Desktop reset the whole connection, reloaded the GPU and renderer, kept about two CPU cores busy, overheated the laptop, and fell into a reconnect loop. Wi-Fi drops or laptop sleep can also break the SSH tunnel and make the same loop happen again.
Memory OS is an open-source tool for adding lasting memory to Hermes Agent. It uses Qdrant to store persistent memory, so past work and useful details can be retrieved later. It can keep structured facts, bring back related information, and build an automatically maintained wiki. It also supports context injection, which means only the needed memory is placed into the agent’s current work. It runs locally and can work with any LLM, so it is not locked to one model provider.
Hermes Agent appears to be built around a small core. That core receives messages, runs the agent loop, and keeps track of each conversation. When a conversation becomes long, the system can compress it so it stays usable. Extra abilities, such as image generation, Google Meet, and messaging, are split into separate plugin packages instead of being mixed into the core. The model layer is also treated like a plugin, so changing the language model does not have to reshape the core system. One backend can connect to several front ends, including a desktop app, a web app, and a terminal interface. Adapter and gateway layers sit between the agent and whatever is talking to it, which helps the project grow without making the core too large.
Hermes Agent 0.17.0 has confusing Docker setup guidance because the official docs and the official GitHub repository show different patterns. The Docker guide uses `docker run` with the `HERMES_DASHBOARD=1` environment variable to turn on the dashboard inside the same container. That looks like a single-service setup where the dashboard and gateway run together. The GitHub repository’s `docker-compose.yml` instead defines two separate services: `gateway` and `dashboard`. That points to a two-container setup, where each part runs separately. Some third-party guides also recommend the split setup because it gives better isolation. The practical question is which setup is the current best practice for Hermes Agent 0.17.0.
A lack of Godot knowledge can waste a lot of time even when AI is helping. Checking results is harder when the project runs without a normal screen, and AI has trouble finding and fixing 3D motion sickness problems well. When progress reports and results were not reviewed closely, and almost every recommendation was accepted, the AI moved in a direction that did not match the intended goal. The built work had to be deleted twice. The work became enjoyable again only after spending time learning more and seeing small parts actually work. Game creation needs strong systems, not only code structure. In Godot, a data-driven design felt easier to work with than leaning heavily on classes, inheritance, abstraction, and interfaces. The performance impact still needs more study, but data-driven design became a core rule, and staying open to multiple models and tools, including Hermes, was useful.
Hermes can be used like a private server where a personal AI agent runs. The setup is meant to go beyond a normal chatbot by letting the agent communicate through different channels and handle real tasks. The agent can be connected to WhatsApp, Discord, Telegram, and similar messaging apps. It can also check files on a computer, receive email in its own inbox, and reply to those emails. Voice chat can be added, along with the ability to receive images or videos and understand what is in them. The main pattern is to place the agent on Hermes, then connect messaging, email, file access, voice, and visual input step by step.
Hermes Agent may not be able to read, rename, delete, or move documents in Craft. After the agent creates a document, it may still be unable to edit that document directly. The working action appears to be adding more blocks to the document instead of changing existing content. It is unclear whether this is normal behavior, a Craft limitation, or a setup problem. Anyone planning to use Hermes Agent for Craft document cleanup or management should test the exact actions first.
Qoren is a service for creating and running Hermes and OpenClaw agents without managing servers or infrastructure. Each agent can run in its own separate environment, or multiple agents can run in the same environment so they can communicate with each other. It includes more than 20 ready-made agent templates for a faster start. Custom agents can be built by changing settings and soul.md in the dashboard, or by chatting with an AI to design the agent together. BYOK is supported, but every plan also includes some built-in usage. Without BYOK, usage stops at the available credit limit, and those credits do not expire.
Hermes Agent’s built-in memory can handle small facts, but it may not be enough for searching a larger personal knowledge base. A gbrain MCP server was connected to Hermes Agent to add semantic search over a full Obsidian vault. The setup runs on Postgres 16 with pgvector and searches 155 pages split into more than 3,000 content chunks. When Hermes Agent actually runs the search, it finds the right background, such as context from a project mentioned three weeks earlier, without needing the details explained again. Hermes registered 92 MCP tools, but those tools were not being added to live chat sessions because of a known issue. A command-line wrapper script worked as a workaround because it could call gbrain directly when needed. The main weakness is reliability: even with a firm instruction to search memory before answering when a person, project, or recurring topic appears, Hermes Agent followed that rule only about 60% of the time and skipped memory search the rest of the time.