Verdict: If you run a Hermes Agent in a Slack, Telegram, WhatsApp, or Discord group and your users roll their eyes every time it fires, the open-source Humalike Hermes plugin is the single most targeted fix available in August 2026. One command adds four behavior layers — turn-taking, persona, theory of mind, and social learning — so the agent learns when to speak, how to say it, and who it sounds like. It shipped #1 on Product Hunt on July 22, 2026, the code is MIT-licensed on GitHub, and new accounts get $20 in free API credits to feel the difference before committing. The plugin is genuinely install-and-go for low-stakes rooms, but it intersects with Hermes internals enough that skipping the update step silently breaks turn-taking — that order matters, and we verify it below.
What is the Humalike Hermes plugin?
The Humalike Hermes plugin is a community-maintained, open-source extension for the Hermes Agent framework by Nous Research that gives a running agent four pieces of social-intelligence behavior, all powered by Humalike's hosted behavioral APIs. Humalike itself is a behavioral infrastructure company — it builds seven behavioral APIs in total, and the Hermes plugin bundles the four most useful: turn-taking, persona, theory of mind, and social learning. (Humalike website, GitHub plugin repo)
The problem the plugin targets is specific: agent builders drop a capable Hermes Agent into a group chat, and within two messages everyone knows it is a bot. It answers every single message. It talks over people. It sends ten lines when one would do. The model is smart — it just does not fit in the room. The Humalike plugin is the social behavior layer that sits beneath the model and shapes the timing, voice, reception-awareness, and group-norm adaptation of each reply.
What does the Humalike Hermes plugin change about your agent?
The plugin changes four concrete behaviors rather than swapping your underlying model. The official Humalike × Hermes page documents all four (humalike.ai/hermes):
| Behavior | What it does | Human analogy |
|---|---|---|
| Turn-taking (flagship) | Reads the chat's rhythm and decides whether to jump in or stay silent; naturalizes the reply so it reads like a person typed it | Knowing when a lull in conversation is yours to fill vs. when someone else is about to speak |
| Persona | /soul enhance turns a one-line description of your agent into a fully-voiced character with a consistent voice |
Hiring an intern and giving them a name, tone, and point of view — not a blank template |
| Theory of mind | Weighs how a message will land for the recipient before sending, and adjusts | "Will this read as condescending if I say it that way to a junior member?" |
| Social learning | Picks up the group's slang, formatting, and level of formality over time | Returning to a workplace after a few weeks and noticing you've absorbed the team's shorthand |
The other three Humalike behavioral APIs — Norms (inside jokes and hidden group rules), Social Memory (long-term recall of who's who across conversations), Social Signals (typing pauses, removed reactions, edited messages), and Social Observability (an insight dashboard on who likes the agent, who's bored, who's annoyed) — are available via the Humalike API surface for builders who want to wire them directly. The Hermes plugin focuses on the four lowest-friction behaviors so installation is one command. (Humalike API overview)
What the plugin deliberately does not do: replace the model. It does not add a new language model — your Hermes Agent keeps its existing provider and model config. The Humalike APIs receive the upcoming reply, shape it, and hand it back. That is why the install is one command instead of a stack migration.
How do you install the Humalike Hermes plugin?
Installation is three steps. The order is critical because the turn-taking behavior hooks into Hermes internals.
Step 1 — Update Hermes first. The plugin patches into Hermes internals; an out-of-date Hermes can leave turn-taking silently unhooked, and you'll wonder why nothing changed.
hermes update
Step 2 — Clone and enable the plugin.
git clone https://github.com/Humalike/hermes-humalike-plugin ~/.hermes/plugins/humalike
hermes plugins enable humalike
Note that hermes plugins enable humalike only records the plugin. Nothing happens until Hermes actually starts — the first-start setup runs when the bot boots and applies every required config change itself, printing what it changed, where, and why.
Step 3 — Start Hermes and approve the login link.
hermes
On first start the plugin prints a login URL — open it in a browser from any device and approve. The Humalike API key lands in ~/.hermes/.env and goes live without a restart. Alternatively, you can send /connect in chat to link an account.
Free start: Every new Humalike account receives $20 in free credits with no credit card required — enough to wire the APIs into an agent and ship a real test before paying for anything. (Humalike API)
What does the plugin auto-configure on first start?
The plugin applies several settings so the agent reads as human rather than as a bot announcement system. The full set is in the GitHub README; the load-bearing ones:
| Setting (file) | Set To | Why |
|---|---|---|
streaming (config.yaml) |
false |
Plugin rewrites the final reply into human-style messages; if Hermes streams the raw draft over it, the naturalization layer is invisible |
group_sessions_per_user (config.yaml) |
false |
Everyone shares one conversation; the bot "sees" the whole room instead of a separate thread per person |
display.tool_progress (config.yaml) |
off |
Hides the "Browsing…/Clicking…" tool chatter so replies read as human |
display.busy_ack_enabled (config.yaml) |
false |
Suppresses "⚡ Interrupting current task…" acks — a real person does not announce that |
display.memory_notifications (config.yaml) |
off |
Hides "💾 Self-improvement review…" posts |
slack.reply_in_thread (config.yaml) |
false |
Keeps one shared conversation per channel instead of starting a fresh thread per message |
Restart the gateway once after the first boot so the new values actually load.
What manual setup can't be skipped?
The plugin automates a lot, but three things still need human hands:
- Persona file location. The plugin reads from
~/.hermes/SOUL.md. If yours lives elsewhere, setturn_taking.soul_path: "/path/to/SOUL.md"inconfig.yaml. - Telegram group privacy. Disable privacy mode in @BotFather and add the group's chat ID to
TELEGRAM_GROUP_ALLOWED_CHATS. - Personal WhatsApp number risk. "Respond to everyone" means the bot replies to literally everyone who messages that number. If the number is personal, set
WHATSAPP_ALLOW_ALL_USERS=falseorWHATSAPP_GROUP_POLICY=allowlist.
For Discord, the README walks through enabling Message Content Intent and Server Members Intent in the Developer Portal, then inviting the bot with a prescribed permission integer. Without the Message Content Intent, the turn-taking gate never sees the channel messages.
What does /soul enhance actually do?
/soul enhance is the persona command. When you send it to a running agent, the plugin reads the existing SOUL.md file, sends it to the Humalike Persons API, and writes back an enhanced version that fleshes out the one-line seed description into a voiced character — backed up to SOUL.md.bak so the original is recoverable. It runs once automatically on first startup (disable with soul_auto_enhance: false) and can be re-triggered manually at any time.
The skill is: spend two minutes on the seed. A one-line "helpful assistant" gives you a flat agent that still reads as generic. A two-sentence seed — "Friendly, straight-talking engineer who quotes 90s consoles and calls out hype. Patient with beginners, terse with pros." — gives you a character people actually like in the room.
Where does the Humalike plugin actually help?
The clearest wins are group settings where a task-capable agent has been ruining the social fabric of the chat. Independent community catalogs broadly confirm the positioning:
- Community and support channels — instead of replying to every "hi" and "thanks" with ten lines of FAQ, the agent stays quiet, then answers once cleanly when an actual question arrives. Read more on running a small-business automation stack with AI agents.
- Team Slack channels — turn-taking paces the agent so it doesn't spam during live meetings or interrupt a senior member mid-thought. See how to build an AI agent team with Hermes Agent.
- Creator communities — social learning lets the agent pick up the local slang over weeks and start sounding like a regular, hyping people up and keeping energy while you're offline.
- Customer support rooms — theory of mind means the agent adjusts to a frustrated customer vs. a curious one, instead of sending the same canned FAQ to both.
For solo one-on-one chats the timing problem is trivial — the second you've got four or five people in a room, the agent has to decide whether to speak at all. That is the hard problem Humalike actually solves.
What is the licensing and pricing model?
The plugin itself is MIT-licensed and open source on GitHub — you can read every line of what it patches into Hermes. (GitHub repo)
The Humalike APIs that the plugin calls are hosted and metered. Each new account starts with $20 in free credits and no credit card required; payments begin only when you wire the APIs into something real. A complete public numeric price schedule beyond those starter credits was not published at launch, so builders should check the current offer before budgeting a production deployment. (Humalike API, Kingy AI launch tracker)
What security provenance and research back Humalike?
Three pieces matter for due diligence before you ship an agent into a group:
- Open research. Humalike publishes academic work. "Designing a Deceptively Human AI Facilitator for Group Chats" and "LoSoNA: A Benchmark for Local Social Norm Adaptation in Group Conversations" are both on arXiv. (arxiv.org/abs/2511.17315, arxiv.org/abs/2606.14600)
- Security attestations (in progress). SOC 2 Type II, ISO 27001, and GDPR compliance are "in progress," not complete. Treat that as vendor claim until the audits finish; do not assume certified coverage today. (Humalike website)
- Open-source plugin. MIT-licensed, 176 stars, 15 forks, 112 commits as of the repo created June 18, 2026. The first-start config changes are printed to your terminal so you can see every edit before you commit to it. Patching into Hermes internals makes updates mandatory — an out-of-date Hermes is the #1 silent break.
How popular is the plugin and who runs it?
The Humalike × Hermes launch earned #1 Product of the Day on Product Hunt on July 22, 2026, with 507 upvotes and 95 comments. (Product Hunt chat log, hunted.space dashboard) It also ranked #5 daily and #8 weekly on Launly, which independently republishes the Product Hunt data.
The company is described as a "tiny 🇪🇸 × 🇵🇱 team" (Spain × Poland) that previously shipped AI products at places like Nvidia and Revolut, per public launch commentary. The Product Hunt launch page names co-founder "Martí" as the maker. This is NOT a Y Combinator-backed "Humanlike" company (which is a different HR-recruitment AI product with the similar domain tryhumanlike.com) — verify you're reading the right source before relying on biographical claims. (Crunchbase Humanlike page is a different company.)
What does this mean for you?
If you already run a Hermes Agent in a group of more than two people and the chat tolerates it without groans, you are the exception. If your agent spams, talks over people, or misses tone, this plugin is the single highest-leverage install you can make today. Use a low-stakes test chat first — turn-taking sometimes misfires on noisy rooms and you want to see where before you loose it on customers. And because the plugin is open source, actually read the first-start diff. It is a free lesson in how agent internals work and lets you confirm nothing silently conflicts with your existing gateway setup.
See also our practical walkthroughs on running a multi-agent Hermes fleet, the Hermes Agent v0.19 Quicksilver update, and how to set up AI agents for personal productivity.
FAQ
Q: Is the Humalike Hermes plugin free? A: The plugin code is free and MIT-licensed. The Humalike APIs it calls are metered: every new account gets $20 in free credits with no credit card required, and you only pay once you wire the APIs into a real deployment. A complete public numeric price schedule beyond those starter credits was not published at launch — check the current offer at humalike.ai before budgeting.
Q: Does the Humalike plugin work in group chats? A: Yes — the plugin is explicitly built for group chats (two or more people) on Slack, Telegram, WhatsApp, and Discord. In a two-person chat the timing problem is trivial; once you've got four or five people, the agent has to decide whether to speak at all. That is the problem turn-taking solves.
Q: How do I stop the agent from messaging everyone on my personal WhatsApp number?
A: Set WHATSAPP_ALLOW_ALL_USERS=false or WHATSAPP_GROUP_POLICY=allowlist in ~/.hermes/.env. The plugin opens "respond-to-everyone" by default so the agent can participate in groups without @mention — if the number is personal, close that down before adding the agent.
Q: Does the plugin replace my Hermes Agent's language model? A: No. It adds a behavior layer beneath whatever model your Hermes Agent already uses. The Humalike APIs receive the upcoming reply, shape it (timing, voice, reception-awareness, norms), and hand it back. Your provider, token spending, and model config stay unchanged.
Q: What is the most common reason turn-taking silently stops working?
A: An out-of-date Hermes install. The plugin patches into Hermes internals, so if your Hermes is behind, turn-taking can fail to hook in and you'll wonder why nothing changed. Always run hermes update before enabling the plugin — the Humalike GitHub README names this as the #1 thing people skip.
Q: Is Humalike SOC 2 certified yet? A: No — SOC 2 Type II, ISO 27001, and GDPR compliance are "in progress," per the official site. Treat that as a vendor claim, not a completed certification, until the audits publish. This matters before you ship an agent into regulated chats.
Discussion
0 comments