Skip to content
vavio

Can someone else talk to my Hermes, and what can they make it do?

Yes, if your Hermes messaging configuration admits them. An admitted person can ask the agent to use the tools and connected accounts available to that conversation; being a colleague rather than the owner does not automatically make their requests read-only. In a shared chat, the upstream default also lets any admitted member answer a pending approval, even when someone else requested the action.

This guide reviews Hermes v0.23.1, commit 3b42157, on 16 November 2026. These are upstream source and documentation findings. We did not connect a shared chat or test channel setup on a Vavio computer for this article.

Four controls answer different questions

Your question Control to inspect What it does not establish
Who may send the bot a request? Platform user lists, pairing, group/chat grants and allow-all settings Which tools an admitted request can use
Who may use an administrative slash command? allow_admin_from, user_allowed_commands and their group equivalents A restriction on ordinary requests written to the model
Who may answer a pending approval? approvals.approvers plus the relevant requester/admin rules That every action will ask for approval
Whose conversation history is loaded? group_sessions_per_user and thread_sessions_per_user Separate operating-system accounts or private files

If you want someone to ask questions without changing anything, check the available tools and credentials as well as the chat’s membership.

A user allowlist is only one way in

Hermes supports platform-specific lists such as TELEGRAM_ALLOWED_USERS, DISCORD_ALLOWED_USERS and SLACK_ALLOWED_USERS, plus a global GATEWAY_ALLOWED_USERS list. Approved DM pairing is another grant. An unknown sender receiving a pairing code has not yet been admitted: the operator must approve it.

For groups, inspect chat-level grants too. In the pinned gateway source, a matching group/chat allowlist can admit a message before the individual sender list is checked. Telegram’s TELEGRAM_GROUP_ALLOWED_CHATS, for example, grants the chat; it is not a list of named colleagues. Do not assume that leaving someone out of the user list excludes them from a chat you have separately admitted. Adapter-level group policies can also affect which messages reach this check.

Discord can admit configured roles, and platform or global allow-all settings can broaden access. The admission rules vary by platform and configuration; an empty user list alone does not tell you that nobody else can talk to the bot. Review the effective configuration of the bot receiving the message, including any paired users and group grants.

Requiring an @mention controls when a bot responds. It does not turn a permitted sender’s request into a lower-privilege request.

An admitted request can do more than produce text

The upstream Telegram, Discord and Slack toolset defaults include terminal access. Actual reach depends on your enabled tools, connected credentials, execution backend and approval settings. A colleague may therefore be able to ask Hermes to read or change files, use a connected service, or create recurring work when the corresponding tools are available.

Restricting slash commands is useful, but it is a separate control. With the admin/user split configured, regular users get the allowed command names plus /help and /whoami. When the admin list is unset for a scope, the upstream compatibility behavior lets admitted users run every registered command in that scope.

Hiding a command does not remove an equivalent model tool. For example, a person who can ask the model to use cronjob_manage may be able to create, change or stop scheduled jobs. Removing their chat access later does not by itself prove that work they created has been cancelled. Check the stored schedules and their delivery targets separately.

Narrowing approvers does not mean owner-only approval

The upstream default is approvals.approvers: ["*"]: any admitted member can answer a pending approval. You can narrow it in the profile’s config.yaml:

approvals:
  approvers: []

With that setting, the requester who started the turn can still answer, as can the configured admins for that chat scope. Listing specific IDs adds those approvers; it does not remove the requester. If you want the owner to review every colleague’s action, this setting alone does not provide that rule.

A turn combining messages from several people has no single requester, so its narrowed approvals go to listed approvers and admins. The relevant admins come from Hermes’s allow_admin_from for DMs or group_allow_admin_from for groups, rather than a blanket assumption about who administers the external service.

These rules govern messaging approval replies. Host interfaces such as CLI and Desktop answer as the host; Matrix and QQBot have additional platform-specific restrictions. Ordinary commands may run without a prompt, smart approval may decide a flagged command, and auto-approval changes the flow. See which actions Hermes asks you to approve before relying on an approval prompt as your boundary.

Separate histories do not make separate computers

The default group_sessions_per_user: true separates participants’ conversation histories in ordinary groups when user IDs are available. Threads are shared by default: thread_sessions_per_user: false. Setting the group option to true therefore does not make every shared thread private. Adapter modes can force sharing too: Telegram observe-unmentioned mode uses a shared group session, and Slack multi-person DMs share a session.

Replies posted in a group remain visible to people who can read that group. Some adapters can also add channel history to the model’s context. A separate session key is not a filesystem or credential boundary: conversations can still use the same configured tools and resources.

If two people need different access, qualify the tools, credentials and execution isolation for each environment. A separate profile can help organize configuration, but profiles alone do not provide operating-system isolation.

Before inviting a colleague

Review who is admitted, what tools that person can trigger, who can answer their approvals, and where replies or scheduled results will appear. Then test a harmless request from the intended person and an excluded person on the exact platform. Verify the action and access result, rather than treating a model’s assurance as proof.

Vavio hosts your Hermes on a managed computer for $14.99 per month, with model usage billed separately. Its MCP server and CLI let a supported assistant help manage the setup. Hosting keeps the computer away from your laptop; deciding who can direct its agent remains a separate choice. Review the current hosting offer if that is the service you need.

Sources

Written by Vavio with AI assistance and checked against official Hermes v0.23.1 source and documentation. Source review only; no shared-chat or channel-setup trial.

More from the Vavio blog: Hermes Agent guides