Tips and usage for hermes-agent.nousresearch.com
MetaHarness is an NPM package that sits around existing AI tools such as Claude Code, OpenAI Codex, GitHub Copilot, and Hermes. It creates a separate agent harness for each code repository, with task rules, tool policies, scoring, receipts, routing, verifiers, and promotion gates. Each run leaves evidence behind, and each failure becomes a signal for improving the harness. A new improvement is not promoted until it passes tests, cost checks, and safety gates. Supported execution surfaces include Claude Code, OpenAI Codex, GitHub Copilot, OpenCode, OpenClaw, Hermes, pi.dev, RVM, and GitHub Actions. The package passed 100,000 downloads in its first two weeks.
Hound V9 is a free MCP server that runs on your own computer. It lets an AI agent search the web, fetch web pages, crawl sites, get past some Cloudflare blocks, read PDFs, and extract text from scanned PDFs. It does not require an account, an API key, or paid search services, and it is released under the MIT license. Search uses several sources, including DuckDuckGo, Brave, Google, Startpage, Qwant, and Wikipedia, then uses a local ONNX model to reorder results. Page fetching uses a Patchright browser setup aimed at pages that block normal Playwright with Cloudflare security checks. Crawling can follow important pages first, use a sitemap-style mode to map a site in one call, and set limits for page count, crawl depth, and character count.
The Hostinger Hermes Agent install screen is not handling an API key correctly. The screen is a Docker-based setup form for Hermes Agent, with fields for an admin username, admin password, and values such as a Nexos API Key. The entered key seems to include a line break or extra unwanted characters instead of staying as one clean line. The value also appears to start with 200. No confirmed fix is included, so the main point is that API key entry during Hostinger setup may be unreliable or easy to get wrong.
Firsthand use suggests a Hermes agent can work well on the first run but become unreliable in a later session. Even when one agent has one clear job, detailed instructions, and agreed steps for how the work should be done, the next-day result can change sharply. Hermes may create .md skill files that describe the workflow, but a new session may act as if it did not check those files or the memory. The agent may stop following the earlier method and choose its own path instead. The setup uses Hindsight for memory, so the problem may not simply be markdown memory being overwritten.
Hermes runs on an Apple Silicon Mac with a large amount of memory and handles most work for a job-site management web app. The app uses a Next.js front end, a NestJS API, Postgres, Redis, PgBouncer, and Nginx, while Hermes creates tasks, writes code, runs checks, deploys, and keeps documentation current. Human review mostly happens through Telegram. The setup keeps everything under `~/Hermes`, with separate mounted folders for Hermes settings, runtime state, app code, an Obsidian knowledge vault, model auth files, command-line tools, and patch files. Hermes runs in Docker alongside SearXNG for web search, Hindsight for searchable memory, Bitwarden Lite for secrets, and hermes-pi for calling several cloud model providers. Work is split into four profiles: local-admin coordinates Telegram, scheduled jobs, and dispatch; coder does most implementation; planner researches and files tasks; qa-tester checks the app with Playwright. A custom hook runs every 60 seconds to move ready tasks forward, clean up stalled workers, assign one task to each free profile, and send Telegram updates only when something changes. The coding profile tries several model providers in order, moving from Z.ai to opencode, Nous, and then local LM Studio if a provider fails or hits a limit. A skills whitelist keeps each profile from loading every available skill, which makes the instructions shorter and more focused. Several Hermes 0.15.1 Python files are patched to fix practical problems around provider reset timing, skill loading, macOS file locking, SQLite startup crashes, read-only task lists, and false stuck-dispatcher alerts. Memory is split into a short per-profile MEMORY.md for rules that must apply every turn, a larger Obsidian vault for long notes, and Hindsight for pulling useful context when needed. Scheduled jobs handle worker cleanup, board maintenance, post-task audits, redeploy checks, model health checks, daily notes, and database backups, while Gitea stores separate backups for config, app code, and project notes.
Two simple themes are available for people who find the default Hermes Agent dashboard hard to focus on. They are named `boring-dark` and `boring-light`, covering dark mode and light mode. The themes reduce the strong teal look, textured background, edge glow, and decorative fonts, then replace them with plainer colors and normal system-style text. To install them manually, place the theme files in `~/.hermes/dashboard-themes`. After that, set the default with `hermes config set display.dashboard_theme boring-dark`, or choose the theme from the dashboard theme switcher. There is also a GitHub install script, and `boring-light` can be passed to that script if light mode should be the default. Hermes reads these `YAML` theme files without needing a plugin. If a theme file is edited while the dashboard is already open, switch to another theme and back again so the dashboard reloads the change.
Hermes Apollo is presented as a voice layer for Hermes Agent, moving the agent beyond a typed chat box. A person can speak a task instead of writing a prompt, then ask for app building, browser actions, briefings, simple timers, language help, or saved memory. The broader pattern is to treat Hermes as a workflow system, not just a desktop assistant, by connecting memory, tools, browsers, external services, and task routines. Mobile use is being handled through workarounds such as Telegram, Discord voice channels, browser shortcuts on the home screen, always-on mini PCs, and other bridges, since there is no single official mobile app described in the source material. A separate Android app claims to run the Hermes agent locally on the device, with local chat, persistent memory, model switching, a project and file browser, task scheduling, usage details, and tool access. Another approach focuses on easier setup for local devices or a VPS through a web UI, plus a kanban board for daily and weekly tasks.
The local agent fleet uses openclaw together with Hermes. Each agent runs on a tiny PC with almost no computing power of its own. The thinking step runs on a separate RTX 3090 machine using GLM 4.7 Flash. Heavier work runs on larger models on a Jetson Thor AGX with 128 GB of unified memory. Coding uses Qwen3 Coder Next, while specialist jobs such as research and code review by a different model family use separate models. One Hermes agent manages the openclaw agents. Another Hermes agent acts like a secretary or receptionist for the messaging contact point. The open question is whether better model choices or role assignments would make this setup more effective.
A local agent setup for research is using models such as Hermes, Clawd, and Odysseus, but weaker local LLMs are still a clear limit. Access to a more powerful computer may reduce that performance problem. The main interest is the agent-style interface seen in services like arena.ai and scispace.com. The desired setup can run Bash scripts, keep files separated for each chat, install Python, JavaScript, and npm libraries when needed, and show outputs through an artifact viewer. The goal is to keep privacy-sensitive research data on the local machine while still getting a smooth agent workflow like modern web AI tools provide.
A planned local Hermes setup uses a 128GB Strix Halo 395+ AI Max PC. The expected usable VRAM is about 96GB, and the main question is which model should act as the everyday driver for Hermes. Hermes 4.3 36B and Qwen 3.6 27B are the two model choices under consideration. The setup may also add an Nvidia 4090 graphics card over USB4, so smaller models or Stable Diffusion can run fully offloaded from the main AMD system. The uncertain part is whether mixing an AMD setup with an external Nvidia card will be difficult or unstable.
The Hermes Agent Accelerated Business Hackathon is for building agents on Hermes that can earn money, spend money, and run real work. The target is business tooling, ranging from a fully automated company to a framework that speeds up enterprise tasks. New Stripe Skills let a Hermes agent buy what it needs, set up its own SaaS tools, and pay for services it uses. NVIDIA integrations let teams run agents more safely through NemoClaw and build faster with NVIDIA agent skills. First prize is $10,000 cash, an NVIDIA DGX Spark, and $5,000 in Stripe credits. Second prize is $5,000 cash, an NVIDIA DGX Spark, and $3,000 in Stripe credits, while third prize is $2,500 cash, an NVIDIA DGX Spark, and $1,000 in Stripe credits. To enter, builders need to post a 1-to-3-minute demo video on X tagging NousResearch with a short writeup, drop the link in the submission Discord channel, and complete the Typeform submission form. Nous Research, NVIDIA, and Stripe will judge entries on usefulness, viability, and presentation, with submissions due by the end of Tuesday, June 30.
Hermes can automatically attach Claude credentials and show Claude agents as available agents. It is not clear from this item whether there is a setting to stop that automatic attachment. Deleting the Claude connection inside Hermes can also sign the person out of Claude Desktop and the Claude command line tool. That means removing the connection in Hermes may affect other Claude tools on the same machine.
Hermes was used to create automation tools on Apify and sell them as API services. The reported revenue was $71.36 in June and $6.84 in the first three days of July. The work was described as having almost no maintenance and a 99.66% profit margin. The suggested process is to have Hermes learn Apify, then create multiple actors. Researching demand is better, but starting with simple ideas is also possible. Only a small share of actors may bring in most of the money, so the approach is to make many, let them run for a few months, keep the profitable ones, remove the rest, and repeat. The practical point is to use Hermes as a tool for creating small sellable automations instead of only spending tokens on experiments.
Since Hermes Agent 0.17, web extract with a self-hosted Firecrawl setup has been failing every time. Hermes Agent and Firecrawl run on the same desktop and the same Docker Desktop instance, but they are in separate containers. After Firecrawl is set as the provider for web extract, requests made from the agent chat always fail. The same scrape request works when it is run as a Python script from inside the Hermes container through the VS Code terminal. The Hermes service runs with the `gateway run` command, exposes port 8642 for the core API gateway and port 9119 for the web dashboard, and enables settings for the dashboard, trusted proxy, config path, and a longer cron timeout.
A local Hermes Agent can be used with an API so a PDF is sent to a computer by email and returned as an audio file. This is useful for turning research material into audiobook-style audio for driving or doing chores. When there is plenty of time to wait, audio quality matters more than fast generation. A larger model or voice clones with longer context could likely improve the listening result. Raising steps does not automatically improve quality and can make the output fit the context too tightly. CFG and Temp changes may make the voice slightly better or slightly worse, but they do not seem to solve the main quality problem.
A self-hosted Hermes agent can be used without taking out a phone by connecting it to Even Realities G2 smart glasses. A small app lets the wearer tap the glasses or ring touchpad, speak, and send the spoken words to the agent through live speech-to-text. The answer streams back onto the glasses display. The display can show the agent’s work process and the tools it is using, such as web search, code, and memory. Sensitive actions can be approved from the glasses with a swipe. Longer tasks can be started and left running in the background. A session list shows whether each run is still active, waiting for approval, finished, or failed, and reopening a session reconnects to the live stream. Opening a session also shows the latest part of the conversation so the user can see the agent’s current context.
Hermes Agent and OpenClaw show that agentic AI can use tools, remember information, run multi-step workflows, handle messages, and perform automation. The hard part is not only capability. Most people do not want to set up an agent framework by hand. They want AI that helps with everyday tasks in a way that feels clear and safe. UI/UX becomes a core part of whether these systems are usable. Wider adoption will depend on visible steps, user approvals, memory controls, and interfaces that make powerful actions easy to understand. Row-Bot is presented as one example of trying to make agentic AI easier for normal users to operate.
Hermes can strongly refuse tasks around personal media server tools such as Radarr, Sonarr, and Jellyfin. It may refuse to search, log in, or add movies to a list. The likely reason is that these tools could be used for copyright infringement. When pushed, Hermes can answer in a judgmental way about piracy or stop helping with the task. The same problem appeared after trying local models, less restricted local models, ChatGPT Codex through OAuth, Deepseek, and other options, so the limit may not come from only one connected language model. ChatGPT can still help troubleshoot Radarr and Sonarr, and OpenClaw handled the same kind of management smoothly in the past. Hermes also appears to resist roleplay as fictional characters, such as a JARVIS-style assistant.
In firsthand use, Hermes Agent handled task delegation, conversation compression, and session management more cleanly than OpenClaw. The setup ran local inference on an Apple silicon M4 Pro with 24GB of memory using oMLX, while a UGREEN DXP4800 Pro network storage device ran Docker with separate containers for Hermes Agent and SearXNG. Telegram became the main client, with WhatsApp used earlier. About 20 quantized local model setups were tested, including Qwen3.5-9B, Gemma 4 26B-A4B, Gemma 4 12B, gpt-oss-20b, and Qwen3.6-35B-A3B. For one request at a time, Gemma 4 26B was faster than Qwen 9B, at about 64 tokens per second versus 38. Qwen only pulled ahead when batching multiple requests, so the main limit was not Hermes itself but the local inference model behind it.
Agent spending can drop through configuration changes without switching the model, framework, or workflow. In one tracked OpenClaw case, about 90% of the token spend came from extra material around the real task, such as background overhead, old conversation history, and unused tool schemas. An agent often sends the whole prior conversation again with each new request, so long sessions make the same old text billable again and again. By the time a session reaches 40 messages, each new answer can include all 40 previous messages as input tokens. The main visible setting is a conversation-history cap of 20 messages. In OpenClaw, the example uses `maxHistoryMessages` set to 20 in `~/.openclaw/openclaw.json`; in Hermes, look for the equivalent history or context limit in the agent config.
An M1 128GB setup was used to test whether local AI models could replace advanced paid models for many coding tasks, but the results fell short. The first goal was to replace about 90% of advanced model work with local coding models, but that proved unrealistic. A second approach used an advanced model for planning and a local model for execution, but that also did not work reliably. Several coding tools were tried, including opencode, openhands, picode, and Hermes. The breaking point was a Hermes kanban board task where Qwen 27B spent millions of tokens trying to fix a table div and still failed. Even with Q8 and long context, many requests reached about 100,000 to 120,000 tokens, while the model still lost context and made unwanted changes. Local models worked better for one-shot jobs like summarization and classification than for long coding-agent workflows.
Hermes works better as a coordinator than as a fully automatic replacement for people. The more practical setup splits work into three groups: tasks a person should handle, tasks AI can usually handle safely, and tasks AI can do but should pause for review before completion. That middle review layer reduces the need to constantly watch the agent while still getting useful automation. In this setup, Hermes runs on a cheaper model such as DeepSeek V4 Flash and manages the flow, while Claude Code does the heavier writing or coding work. A Kanban board sits between them so each task has a visible state such as planned, active, review, or done. Wrapping AI calls in deterministic scripts helps prevent missed status updates. Hosted Hermes services may already include identity, memory, and gateway setup, while self-hosted setups may require files such as SOUL.md to be configured by hand.
SBET’s ETH concentration fell from 4.04 to 3.94, and a Hermes agent was used to check whether this was real NAV dilution or something else. From June 22 to June 28, 2026, SBET raised $75.0 million by issuing 10,013,351 new shares and about 10.01 million warrants at $7.49 per share, with $73.5 million coming in after costs. During the same period, it spent $16.1 million to buy 10,000 ETH at an average price of $1,611, and spent $10.0 million buying back 2,132,773 shares at an average price of $4.69. ETH concentration means total ETH holdings divided by shares outstanding. The share count rose 3.63% on a net basis, or about 7.88 million shares, while treasury ETH rose only 1.14%, or 10,000 ETH. The drop happened because only $26.1 million of the new cash had been used so far, while $47.4 million was still sitting as cash on the balance sheet.
A firsthand case shows Claude subscription limits dropping faster than expected when Claude is connected through Hermes. The problem appeared on the day Fable opened. The same Claude subscription still worked normally in T3 Code, another execution tool. That points away from the Claude subscription itself and toward a possible Hermes connection issue, especially OAuth.
A small Hermes plugin is being tested to choose tool groups before the main answer is generated. The goal is to avoid loading every tool schema into the first request when the task does not need them. Simple prompts can start with almost no tools. URL or research tasks can load only browser and web tools, while coding or local file tasks can load file and terminal tools. A recovery tool stays available in case the router chooses too narrowly. If the router is unsure, it falls back to the full tool list instead of blocking the task. In one test profile, a prompt containing only a period was reduced to almost just the recovery tool, while browser and research prompts still loaded the right web tools. The first request was about 4.25k tokens. This is still a test plugin, not polished public software.
OpenCode 2.0 beta is expected to make it easier to create skills while working away from a full setup. That could reduce the number of prompts needed when using Hermes. OpenCode’s hot reloading fits the same practical need: change a skill, test it quickly, and keep improving it without slow restarts. The TUI is light enough to be included inside a Hermes workflow. There is also a working setup using Hermes with OpenCode on a Lightnode VPS. Better SSH support would make this kind of remote workflow easier and more reliable.
Agent memory sounds like one feature, but it is really a set of different methods that do different jobs. Most real agents do not need every kind of memory; they usually use only two or three that fit the task. One important kind is working memory, which means the information currently inside the model’s context window. This matters in long chats, coding sessions, or multi-step work where the agent must keep track of the current goal, recent changes, and next steps. The open-source project Letta treats the context window like RAM: it keeps the most important details close by and moves older information into searchable storage. The tradeoff is that the context window has limits, and more tokens can raise cost.
AI models and agent tools are becoming similar fast, so the lasting value is not the tool itself but the context layer: personal research, notes, and domain knowledge. Skills, agents, and workflow logic built inside Claude Code can become locked to that ecosystem. Moving to another open-source tool such as Pi, Hermes, or OpenCode does not solve the problem if the context layer is still not portable. Real freedom comes from pulling memory out of the tool and designing the context layer separately. A context layer can be split into three parts: unified memory, business logic, and a serving layer. With that setup, the tool on top becomes replaceable. The simplest starting point is a file-based system, then one database that can handle text, vector, and graph search if needed. Multiple specialized databases should only come later if they are truly necessary.
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.