Hermes Agent skills for social media: post, reply and schedule from your agent

Hermes Agent ships an X skill and nothing for Instagram, LinkedIn or TikTok. Skills are markdown procedures, plugins are Python, and neither one holds your accounts. An MCP server does: add PlugKit and the agent gets over 100 tools to publish, read comments and answer DMs. Then a skill turns them into your routine. Facts checked September 25, 2026.

Skills, plugins and MCP servers: three layers people mix up

The three words appear in the same sentence on every Hermes thread, and they are not interchangeable. Straight from the Hermes docs on skills, plugins and MCP, read on September 25, 2026:

A skill is procedural memory. It loads in three levels to keep the context small: skills_list() returns names and descriptions for about 3k tokens, skill_view(name) opens one, skill_view(name, path) pulls a single reference file. It tells the agent how to work. It holds no token itself: it can ask for an API key kept in ~/.hermes/.env and ship helper scripts, but the account connection lives elsewhere.

A plugin is real code. hermes plugins install owner/repo runs a static security scan first, with three verdicts: safe installs, caution asks for confirmation, dangerous is blocked. Tools are registered with ctx.register_tool(...) and show up next to the built-ins once the plugin is enabled (hermes plugins enable <name>) and Hermes restarts. You could write an Instagram plugin. You would also be writing OAuth, token refresh, media containers and Meta app review yourself.

An MCP server is the shortcut. Hermes has a native client, tools are namespaced mcp__<server>__<tool> (double underscores since v0.19.0), and they are available in every surface: CLI, desktop, Telegram, Discord, cron jobs.

The Skills Hub has no skill that holds your accounts

hermes skills browse and hermes skills search <query> open the hub, hermes skills install <source/skill> adds one, hermes skills tap add <owner/repo> subscribes to a repository of your own, and /learn <source> writes a skill from a URL or a procedure you describe. It is a large catalog: its index listed over 100,000 skills on September 25, 2026.

What you will not find there is a skill that holds an Instagram account for you, and that is not an oversight. Publishing needs a business or creator account, a reviewed Meta app, an access token and a container flow. The community skills that do post to Instagram ask for your own Meta app and token, or hand the job to a third-party service that holds the account. Hermes itself bundles one social skill, xurl, which posts to X through X's own CLI and your own X developer app. TikTok's Content Posting API, YouTube's Data API and LinkedIn each add their own authorization and review.

So the honest order of operations is: plug the accounts in first, write the skill second.

Step 1: add PlugKit as an MCP server

PlugKit is a hosted MCP server at https://api.plugkit.co/mcp over Streamable HTTP, so there is nothing to run locally. In Hermes:

hermes mcp add plugkit --url https://api.plugkit.co/mcp --auth header
hermes mcp test plugkit

The key is written to ~/.hermes/.env and referenced from the config, never inlined:

mcp_servers:
  plugkit:
    url: https://api.plugkit.co/mcp
    headers:
      Authorization: Bearer ${MCP_PLUGKIT_API_KEY}
    enabled: true

hermes mcp test <name> verifies connectivity and reports the tool count. For the full walkthrough with the real terminal output, the startup quirk that makes the first message look broken, and the OAuth alternative, read Hermes Agent setup: connect Instagram, LinkedIn and X.

What the agent can do once the accounts are in

The PlugKit MCP docs list over 100 tools and 3 workflows as of September 25, 2026. Grouped, with the names your agent will actually call:

Three workflows ship as prompts: publish, inbox, report. Clients surface them as commands.

The important detail for an always-on agent: a session granted the mcp:read scope never sees the write tools at all. They are hidden, not refused. Scopes come with the OAuth sign-in: the API key from Step 1 has no scope and carries full access. Run a week of reporting on read-only OAuth before you hand over publishing. On the Hermes side you can narrow further with tool filters:

    tools:
      include: [whoami, list_posts, get_analytics_summary, list_engagement]

Write the skill that turns tools into your routine

Now the skill is worth writing, because there is something for it to drive. The frontmatter fields come from the Hermes docs; keep the description under 60 characters, or the skill index cuts it off:

---
name: weekly-social-report
description: Pull last week numbers and draft Monday post
version: 1.0.0
metadata:
  hermes:
    tags: [social, reporting]
    category: marketing
---

The body uses the four sections Hermes expects: When to Use, Procedure, Pitfalls, Verification. Mine reads roughly like this:

1. Call mcp__plugkit__whoami, take the default profile.
2. Call mcp__plugkit__get_analytics_by_account over 7 days, one row per account.
3. Call mcp__plugkit__get_top_posts ranked by reach, keep 3.
4. Call mcp__plugkit__get_best_times_to_post, pick Monday's best hour.
5. Draft the post, show it to me, and only then call mcp__plugkit__create_post with scheduledFor.

Pitfalls worth writing down: the analytics tools read a collected store and cost nothing, while get_account_analytics and get_post_analytics hit the platform live and are billed on X. Two accounts on the same platform need get_analytics_by_account, not the summary. And send_message is immediate and irreversible, so a draft step before it is not optional.

Skills stack in chat: /weekly-social-report /brand-voice go loads both.

Schedule it with Hermes cron

hermes cron create "<schedule>" "<prompt>" registers a job, hermes cron list, pause, run and runs manage it. Jobs live in ~/.hermes/cron/jobs.json, outputs in ~/.hermes/cron/output/, and the gateway ticks the scheduler every 60 seconds, running each due job in an isolated session.

One detail: a cron job runs with the toolsets set for the cron platform in hermes tools. Enabled MCP servers come along by default, but if that list names other MCP servers or no_mcp, add plugkit to it, or the job will report that it cannot post and you will blame the MCP server.

Output can be delivered to Telegram, Discord, Slack, email or a file, which is how a Monday report reaches you without you opening a terminal.

What the whole stack costs

Every account starts on a 7-day trial. X background work is off by default and every account carries a $20 monthly spend cap until you set your own, which matters when an agent is the one deciding to poll. Prices read on docs.plugkit.co on September 25, 2026.

If you are choosing your agent runtime rather than your plug, OpenClaw vs Hermes Agent for social media compares the two on exactly this job. If you are on Claude instead, the Instagram MCP server page covers the same tools from Claude Code and Cursor, and the agents hub lists the rest.

PlugKit is $29 a month for 5 connected accounts, cancel anytime.

FAQ

Does Hermes Agent have a social media skill?

For X, yes: the bundled xurl skill posts through X's own CLI, with your own X developer app. For Instagram, TikTok or LinkedIn, nothing built in. The Skills Hub has community skills for them, but each one needs your own Meta app and token or a third-party service holding the account. You give the agent accounts through an MCP server or a plugin, then a skill orchestrates them.

What is the difference between a Hermes skill and an MCP server?

A skill is knowledge: instructions the agent loads when relevant, stored in ~/.hermes/skills/. An MCP server is capability: an external service whose tools Hermes discovers at startup and registers as mcp__<server>__<tool>. Skills decide what to do, MCP tools do it.

Do I need a plugin to post to Instagram from Hermes Agent?

No, and writing one is more work than it looks. A plugin can register a post_to_instagram tool, but you own the Meta app review, the OAuth refresh, the media container flow and the rate limits, times every platform you add. An MCP server that already holds the connections skips all of it.

Can I stop the agent from publishing by mistake?

Yes, at three levels. Connect through OAuth and grant only mcp:read: the write tools are not exposed at all (an API key has no scopes). Use tools.include in ~/.hermes/config.yaml to whitelist a short list. Set trust: untrusted on the PlugKit server so every tool not marked read-only waits for your approval, and put a confirmation step in the skill itself before any create_post or send_message.

How many tools does the PlugKit MCP server expose to Hermes?

Over 100 tools and 3 workflow prompts, per the PlugKit docs on September 25, 2026, covering publishing, the unified inbox, comments, analytics, comment-to-DM campaigns, Meta ads, Instagram shopping and billing. Your own count depends on the scope you grant, since write tools are hidden from a read-only session.

AB

Antoine BlancoFounder of PlugKit. Runs his own Instagram, YouTube and X accounts through the API and MCP server described here.