Nous Research shipped the Hermes Agent desktop app on June 2, 2026, and the project has now crossed 180,000 GitHub stars in under four months since its February launch - making it the fastest-growing open-source agent framework of 2026 by Dealroom's count. Two days ago, v0.18.0 dropped on July 1, clearing every P0 and P1 issue in the entire repo - roughly 700 highest-priority items - leaving zero open P0s and zero open P1s. That is an unusual milestone for a project this young, and it matters for teams considering production use.
The reason Hermes Agent is worth paying attention to is not its GitHub star velocity. It is the architecture. Most AI bots in Slack are thin wrappers: a webhook receives a message, the message goes to an API, a reply comes back. The session ends. The next message starts fresh. Most assistants still operate as session tools - you open them, feed context, finish the task, and lose part of that context in the next session. Hermes is designed around the opposite assumption.
What a self-improving agent actually does differently
Hermes Agent is an open-source (MIT) autonomous agent runtime: a self-hosted, always-on daemon that remembers across restarts, writes and reuses its own skills, schedules its own jobs, and acts through a sandboxed terminal and file system. The skill system is the part worth understanding concretely. Instead of asking the model to solve each task from scratch, Hermes can convert a solved workflow into a reusable skill. The next time a similar task appears, the agent can load that skill, apply prior steps, and refine it.
In practice, that means if someone in Slack asks the agent to pull last week's deploy log, parse the error counts by service, and post a summary to a channel - and the agent figures out how to do it - it saves that workflow. The next engineer who asks a similar question gets a faster, cheaper answer because the model does not have to reason from scratch.
Skill creation happens autonomously after complex tasks, and skills self-improve during use. The agent is also running FTS5 session search with LLM summarization for cross-session recall
- so it can answer "what did we decide last Tuesday about the rate limiter?" without being fed a transcript.
This is not magic. It is a specific architectural choice, and it comes with the operational weight of a persistent daemon you have to run and monitor. That tradeoff is real and covered below.
Running Hermes in Slack and Microsoft Teams
The messaging gateway is a single background process that connects Hermes to over 20 platforms, including Slack, Telegram, Discord, WhatsApp, and Microsoft Teams. It handles sessions, runs scheduled automations, and delivers voice messages across all configured platforms simultaneously. Slack gets full feature support: file attachments, threaded replies, emoji reactions, typing indicators, streaming responses, and voice transcription.
The Slack setup is cleaner than most self-hosted bots. You run one command to generate a JSON manifest, paste it into api.slack.com/apps, and the OAuth scopes and event subscriptions are pre-configured.
When someone @mentions the bot in a channel, Hermes replies in a thread. In DMs it responds to every message without needing a mention. The agent's persistent memory means it builds context over time - ask it to research a topic on Monday and it still remembers the findings on Friday.
You can also bind different skill sets to different channels. The research channel gets access to academic search tools, while the sales channel loads CRM-related skills. That kind of per-channel configuration is not something you get out of most managed bot products.
Microsoft Teams is supported but requires slightly more infrastructure work. Unlike Slack's Socket Mode, Teams delivers messages by calling a public HTTPS webhook, so your instance needs a publicly reachable endpoint - either a dev tunnel for local development or a real domain for production. For teams already running internal tooling on a VPS, this is minor. For teams with no self-hosting experience, it is a real setup cost.
The agent is model-agnostic by design. You can use Nous Portal, OpenRouter, OpenAI, your own endpoint, and many others - switch with
hermes modeland there are no code changes, no lock-in.
The cost structure is genuinely different
Commercial agent platforms typically charge $30-$50 per month at the low end, scaling into hundreds for team plans. Self-hosting Hermes costs roughly $5-10/month for a VPS, with LLM API costs around $0.30 per complex task on budget models. The agent software itself is MIT-licensed and free.
That math becomes interesting when you factor in what is happening to inference prices more broadly. GPT-4-equivalent performance now costs roughly $0.40 per million tokens versus $20 in late 2022.
According to the Stanford HAI 2025 AI Index Report, the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024. Running a capable open-weight model against this infrastructure is no longer an exotic cost.
There is a catch. While lower token unit costs will enable more advanced GenAI capabilities, these advancements will drive disproportionately higher token demand. As token consumption rises faster than token costs fall, overall inference costs are expected to increase. An agent with a persistent memory loop and skill-improvement cycles generates more tokens per session than a simple chat wrapper. The per-token price is low; the per-session token count is high. You need to monitor it.
The Hermes 4.3 36B model fits on a two-GPU RTX 4090 setup at 4-bit quantization or a single A100 80GB at higher precision
- useful if you want fully local inference with no external API calls at all.
What this is and what it isn't
Hermes models (Hermes 3, Hermes 4) are open-weight language models. They are not agents. The agent runs on top of an agentic model tuned for fast, repeated tool calls, which is a different job from long-form reasoning. Hermes Agent is a separate, official open-source project: a self-hosted autonomous agent runtime that drives an agentic model. If you have been evaluating the Hermes 4 model weights on benchmarks, you are looking at the foundation, not the product.
The project includes Atropos RL environments, batch trajectory generation, and trajectory compression for training tool-calling models. This gives Hermes a dual role: it is an operational agent and a data generation pipeline for the next generation of agents. That is not a pitch - it is a reason the architecture is more considered than most open-source agent repos that appeared around the same time.
The skill self-improvement loop is the thing to audit carefully before deploying to a team. Hermes includes built-in prompt injection scanning and credential filtering. However, there is currently no explainability interface for the agent's skill retention decisions. Regular audits of the skills directory are recommended. Skill files are human-readable Markdown, which helps - but you should review what the agent is teaching itself, especially in channels where it has write access to external systems.
For teams comfortable running their own infrastructure and willing to do that review, this is the most capable open-source agent in Slack right now. For teams who want to install something and forget about it, a managed product is still the right call. The gap between those two positions has narrowed considerably since February - and the v0.18.0 stability work suggests it is still closing.