An open-source AI agent from a lab best known for fine-tuned language models has become one of the most-discussed tools among AI builders this weekend. Nous Research’s Hermes Agent runs on almost any model — Grok, Claude, GPT, GLM, Kimi or a local model — with the user’s own keys, and keeps its memory, skills and “personality” when the model underneath changes. A detailed, unsponsored review posted on November 8 by the account @Da7_Tech drew hundreds of likes and a direct thank-you from Nous Research. It also came with a list of failure modes that every team considering an open agent should read.
Quick Take
- What it is: Hermes Agent is a free, MIT-licensed AI agent for desktop, terminal and messaging apps. It is model-neutral: you bring a provider key, run a local model, or buy credits through Nous Portal.
- Why it’s trending: Reviewer Da7em called it “a victory for what open source is supposed to mean,” citing about 165 built-in skills, 94.4% cache hits in an average session and stable operation in a 2 GB RAM virtual machine.
- Pricing: The agent itself is free. Optional Nous Portal plans are Free (pay as you go), $20, $100 and $200 a month; paid tiers include one hosted cloud agent.
- The catch: The same review warns that weaker models “can wreck your Hermes in minutes,” that steering messages are often ignored until a task ends, and that computer use breaks under heavy load.
What happened
On November 8, Da7em (@Da7_Tech), who publishes model tests and comparisons on X, posted what he called a “full review” after “using Hermes Agent every day since day one.” He stressed that it was “completely independent. Not paid, not sponsored.” The review compared Hermes with OpenClaw, Claude Code, Codex and Grok Bot, and quickly circulated in English, Spanish and Japanese. Nous Research replied: “Appreciate the thoughtful review!”
The timing fits a busy fortnight for the project. Nous says it released the Hermes Mobile beta at its NousCon 2026 event and that the beta “filled up quickly”; a waitlist is now open for iOS and Android. Both the TestFlight and Android test programs opened on October 31 and were full by November 1, according to a Hermes Mobile beta guide by Sumus. Version 0.22.7, which shipped on November 3, rolled up about 174 pull requests, according to a release summary. And on November 8, Nous announced an official Security plugin that maps attack surfaces, splits code across parallel sub-agents and exports findings to SARIF.
Cofounder Teknium (@Teknium) has been answering users and promoting community plugins on X throughout — including a time-travel checkpoint plugin whose developer thanked him for fast marketplace approval.
Background: a model lab builds an agent
Nous Research made its name with the open “Hermes” family of fine-tuned models. Hermes Agent applies the same open philosophy to the agent layer that companies such as Anthropic (Claude Code), OpenAI (Codex) and SpaceXAI (Grok Bot) keep proprietary. According to the official site, “Hermes Agent is free and open source under the MIT license. Model providers and optional hosted services have their own pricing.” Stopping a local agent “does not delete its saved conversations, memories, or skills.”
That separation is the design idea at the center of the current praise. Proprietary agents tie the product to one vendor’s models, quotas and policy changes. Nous made the point bluntly on November 8: “Using an agent that can change its models or code at any time without telling you is like building on quicksand.” The context is a week of complaints about quota exhaustion and perceived model changes elsewhere, covered in Grok Bot token limits and the Claude Opus 5.5 “nerf” debate.
How it works
Hermes Agent installs as a desktop app (macOS, Windows) or via a terminal script on Linux. The user connects a model provider, then talks to the agent in the desktop app or through messaging channels; Da7em lists iMessage, Telegram, Slack and Discord. Tools cover web search, browser, terminal, files, vision, image generation, sub-agents, scheduled tasks, memory and computer use, with MCP for anything else. Skills and plugins install from a marketplace.
The piece reviewers like most is what Da7em calls “the layer above the model”: identity, memory and skills that survive a model swap. A “journey” view shows memory and skills as an editable map. The review’s average session — out of 1,013 — used 219 million input tokens, 15.7 million output tokens and 3.77 billion cached tokens, “94.4% from cache.” Those figures come from one user’s setup and depend heavily on provider caching.
| Nous Portal tier | Price | What it adds | Cloud agent |
|---|---|---|---|
| Free | $0, pay as you go | Free models only | — |
| Plus | $20 / month | Free and paid models, hosted tools (web search, images, speech, cloud browser) | 1 agent, 4 GB RAM, 25 GB disk |
| Super | $100 / month | 5× Plus usage | 1 agent, 4 GB RAM, 100 GB disk |
| Ultra | $200 / month | 10× Plus usage | 1 agent, 8 GB RAM, 100 GB disk |
Plan details are taken from the Hermes Agent site as of November 9, 2026. A trending summary on X described the plans as “free to $20 a month”; the official table also lists $100 and $200 tiers.
Pros and cons
For users
- Pro: No Hermes-side quota, free with your own keys or local models, and memory that persists across model changes.
- Con: Not the easiest start — Da7em says it “is not the easiest tool to start with” and wants an “easy mode” with ready-made profiles. Mobile is still a waitlisted beta.
For developers
- Pro: Open code, plugin marketplace and fast maintainer response; Da7em says a feature he requested on X — per-task model selection for scheduled jobs — shipped “in less than three hours,” and his Arabic localization pull request was merged.
- Con: Steering is “hit or miss,” sub-agent results can arrive out of order, and long goal-mode tasks occasionally loop or end early, per the review.
For companies
- Pro: Model neutrality reduces vendor lock-in, and self-hosting keeps the agent’s memory and skills under the company’s control. The new Security plugin exports to SARIF for existing tooling.
- Con: Da7em calls weak models “the most dangerous problem”: a hallucinating model can corrupt the memory and identity layer. He asks for an approval gate on edits to memory, the “soul file” and AGENTS.md — a governance feature enterprises would need too. He also praises that Hermes “refuses less than any agent I use,” which some security teams will read as a risk rather than a feature.
What X is saying
I have been using Hermes Agent every day since day one, and I think it is finally time for a full review. As always, this is completely independent. Not paid, not sponsored. The good and the bad, nothing held back.
— Da7em (@Da7_Tech) November 8, 2026 · View on X
Using an agent that can change its models or code at any time without telling you is like building on quicksand
— Nous Research (@NousResearch) November 8, 2026 · View on X
Hermes Agent can now review your codebase for security bugs with the new official Security plugin by @mark_s.
— Nous Research (@NousResearch) November 8, 2026 · View on X
Not everyone is a fan. Developer Chase Brower graded several agents harshly, writing “openclaw: F- / hermes agent: F- / dots: F+ / duet: D” and calling them “absolutely horrible.” In a Japanese recap, @natsume_aurlia highlighted the review’s claims about no Hermes-side cap, 94.4% caching and running on 2 GB of memory as especially appealing for someone frustrated with Grok Bot usage limits.
What Grok says
Summarizing the review, Grok described Hermes Agent as “an open-source, model-neutral AI agent” that “keeps a persistent ‘soul,’ memory and skills layer separate from the LLM, so you can swap models without losing setup,” while noting the reviewer’s “minor issues: occasional UI lag, slow steering, fragile computer-use and weak models can damage the agent.”
Critical analysis
The review is valuable precisely because it is not a launch post: it lists failure modes alongside the praise. But it remains one power user’s experience. The caching and RAM figures are self-reported, the comparison with Grok Bot (“extremely low on intelligence right now”) is an opinion from a user who prefers to customize, and independent benchmarks tell a more mixed story. On Joël Niklaus’s Harness Leaderboard, Hermes Agent is statistically tied for first on average across three open models and is the best harness for DeepSeek-V4.1-Flash, while Garry Tan’s early evals found Hermes in the middle of the pack on cost ($0.24 per solved task on Opus) — see our analysis of the harness tests.
The structural argument is stronger than any single benchmark. An agent whose memory and skills belong to the user — and outlive any one model — gives buyers leverage that closed agents do not. The flip side, as the review shows, is that the user also owns the failure: if a weak model corrupts the memory layer, there is no vendor safeguard unless someone builds one.
What’s still open
- When will Hermes Mobile leave beta and reach the App Store and Google Play?
- Will Nous add approval gates or versioning for memory and identity edits, as reviewers request?
- How will the hosted cloud agent in paid Nous Portal plans compare on reliability with Grok Bot or Perplexity’s always-on agents?
- Will independent evals confirm the token-efficiency claims across providers?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on Da7em’s November 8 review on X, Nous Research’s official Hermes Agent website and announcements, and public posts on X from November 3–9, 2026. Usage, caching and RAM figures are the reviewer’s own and were not independently tested. Related: Perplexity Teammates, Grok Bot agent teams.
i10x Perspective
Hermes Agent’s moment says less about one product than about where users want control. As closed agents tighten quotas and change models without notice, builders are discovering that the valuable asset is not the model but the accumulated memory, skills and preferences around it — and they want to own that layer. Open agents like Hermes make that possible today, at the cost of more setup and more responsibility. The next step that would make them enterprise-ready is boring but essential: versioned, reviewable memory, so that “the layer above the model” can be audited and rolled back like any other critical system.
