What Does the Hermes Agent Actually Do in Slack and Teams?

Hermes has amassed roughly 214,000 stars and nearly 40,000 forks on GitHub

• numbers that tend to attract breathless coverage. Most of it skips the part where you spend a weekend configuring Azure AD. This post covers what the Hermes agent actually is (distinct from the Hermes LLM line), what it does inside Slack and Teams once it's running, and whether the operational cost makes sense for a real team.

What the Hermes agent is, and what it is not

The name causes confusion because Nous Research ships two things called Hermes. The model line - Hermes 4, Hermes 4.3 - is a series of post-trained LLMs. Hermes 4.3 was released August 25, 2025 and is built on ByteDance's Seed 36B base model, trained using Nous Research's Psyche decentralized training network. That is the model.

The Hermes agent is something different: a self-hostable agent runtime that connects to messaging platforms and runs tasks on behalf of a user. It lives where you do - CLI, Telegram, Discord, Slack, WhatsApp, Signal, and 20+ other platforms - and supports scheduled automations, a Bot Mode for specialist bots that work together in group chats, and the ability to spawn isolated subagents for parallel workstreams. You choose which LLM it calls; the Hermes model is one option, not a requirement.

Hermes shipped with built-in skills such as web search, coding, and image understanding, and was designed to automatically learn from usage and build more skills without manual intervention. That self-improvement framing is where the marketing gets ahead of the reality - more on that below.

How the Slack and Teams integrations actually work

Hermes runs as a gateway bot on 27+ messaging platforms, all configured through the same gateway subsystem , which includes both Slack and Microsoft Teams.

Slack is the simpler path. Hermes uses Slack's Socket Mode, which means no public inbound endpoint is required - the connection is outbound from your server to Slack's API. Setup is still non-trivial (you need a Slack app, bot token, and a server to host the runtime), but it follows the standard Slack bot pattern most platform engineers have done before.

Teams is harder. 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.

Hermes v0.14.0 (released May 16, 2026) added native Microsoft Teams support as its 20th messaging platform. The integration includes Microsoft Graph authentication, webhook-based message ingestion, the standard Hermes agent pipeline, and outbound delivery via the Graph API. Setup requires an Azure AD app registration, a public HTTPS endpoint, and explicit user allowlisting.

Time estimate from primary documentation: with Azure experience, expect one to two hours for basic bot setup and testing; without Azure experience, plan for three to five hours.

Once running, the approval model is the genuinely interesting part. When the agent needs to run a potentially dangerous command, it sends an Adaptive Card with four buttons instead of asking you to type /approve. Clicking approves inline. That is closer to a proper draft-and-approve flow than most home-built bot setups manage.

The meeting pipeline is a secondary surface worth noting. When the Teams meeting pipeline plugin is enabled, the adapter also handles outbound delivery of meeting summaries - one Teams integration surface, not two. After a meeting's transcript is summarized, the writer posts the summary into your chosen Teams target.

What the self-improvement claim actually means

The changelog entry that gets the most attention is the autonomous Curator. The release ships an autonomous background Curator that silently keeps your skill library sharp, a substantially upgraded self-improvement loop, four new inference providers, native Microsoft Teams integration, and a 57% reduction in TUI cold start time.

"Self-improvement loop" sounds like the agent is rewriting itself. What it actually does: the Curator monitors which skills are getting used, flags ones that are stale or broken, and can fetch updated versions from the Skills Hub. Skills are compatible with the agentskills.io open standard, portable, shareable, and community-contributed via the Skills Hub. That is genuinely useful - skill rot is a real operational problem in self-hosted agents - but it is closer to a package manager than autonomous improvement.

The multi-agent layer is more substantively new. Hermes v0.18.0 adds first-class Mixture-of-Agents models, completion contracts and verification evidence for /goal, /learn, and /journey, and background fan-out. In practice this means you can define a goal, Hermes breaks it into subtasks, and subagents run them in parallel - with each step producing a logged receipt you can inspect.

The real cost model before you commit

Hermes Agent is free under an MIT license. But the Teams integration requires a server with a public HTTPS endpoint ($10-$50/month VPS), Azure AD (included with Microsoft 365 business plans), and your LLM API costs. Ongoing maintenance includes secret rotation, allowlist management, and webhook renewal.

The LLM API cost is worth computing. If you route most traffic through an open-weight model via OpenRouter - say, Hermes 4 70B at $0.13 per million input tokens and $0.40 per million output tokens

• a team generating 5 million tokens a month (roughly 200 substantive agent interactions a day) spends around $2,650 per year on model costs alone, before the VPS. That is materially cheaper than most closed-model alternatives at equivalent volume.

The non-obvious cost is ops time. Hermes requires allowlist management, secret rotation on a schedule, and someone who knows what to do when the Graph webhook stops delivering. The most common cause of broken Teams delivery is the TEAMS_ALLOWED_USERS list - Hermes silently drops messages from users whose AAD object ID is not in the allowlist. Silent failure is hard to debug in production.

Is the capability good enough for work tasks?

This is the question that actually determines whether any of this is worth the setup. On the model side, since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months in the Epoch Capabilities Index, with an average ECI gap of 8 points - similar to the gap between GPT-5 and GPT-5.5.

Four months of capability lag sounds significant until you consider what most in-channel work tasks actually require: finding a number in a document, summarising a thread, routing a question to the right person, or drafting a reply for review. Open-weight models dominate the low- and mid-cost portion of the cost-performance frontier; the lowest-error end is largely defined by closed-weight models, suggesting proprietary systems disproportionately push peak benchmark performance. For tasks where "good enough and fast" beats "maximally accurate," the open-weight tier wins on economics.

The honest caveat: two factors mean the gap estimate may understate the true distance - open-weight models tend to perform worse on private benchmarks compared to closed models, plausibly because they more aggressively hillclimb on public benchmarks. If your use case depends on consistent performance on your own, private data rather than public benchmarks, budget time to evaluate on your actual workload.

A teammate like Beagle runs on a managed, tested model layer - so you get the draft-and-approve flow without the Azure AD weekend. For teams who want to run their own infrastructure, Hermes is a serious option. For teams who want the capability without the ops overhead, the tradeoff is worth examining before committing.

Hermes agent in Slack and Teams: common questions

What is the Hermes agent, and how is it different from the Hermes model?

The Hermes agent is an open-source agent runtime built by Nous Research. It connects to messaging platforms and executes tasks using whatever LLM you configure. The Hermes model line (Hermes 4, 4.3) is a separate series of post-trained language models. You can run the agent without using a Hermes model at all.

Does Hermes Agent work natively in Microsoft Teams?

Yes. Hermes v0.14.0, released May 16, 2026, added native Teams support via Microsoft Graph webhooks. Setup requires Azure AD app registration, a public HTTPS endpoint, and explicit user allowlisting. Expect one to five hours of setup time depending on Azure experience.

What does a Hermes agent approval flow look like in Teams?

When Hermes needs to take a potentially consequential action, it sends a Teams Adaptive Card with inline approve/reject buttons. Clicking resolves the decision without leaving the thread. The result is logged with a reason automatically.

How much does running Hermes in production actually cost?

The software is MIT-licensed and free. Ongoing costs are a VPS ($10-$50/month), LLM API fees (roughly $0.13-$0.40 per million tokens for open-weight models via OpenRouter), and engineer time for maintenance. Silent failures from misconfigured allowlists and webhook renewal are the most common operational issues.

Are open-weight models good enough for in-channel work tasks?

For most in-channel work tasks - lookups, summarisation, drafting, routing - yes. Epoch AI's 2026 analysis puts the best open-weight models four months behind the closed frontier on aggregate capability. The cost difference (often 10-50x cheaper per token) makes open-weight the practical choice for high-volume, routine work-task automation.