Getting the Most Out of Hermes Agent

A practical field guide to Hermes' superpowers β€” MCP integrations, skills, and keeping your token spend lean. Hermes is an open-source AI agent by Nous Research that runs in your terminal, desktop, and 20+ messaging platforms with full tool access.

Daily Power Moves

The features that separate casual use from genuinely getting the most out of Hermes.

🧠 Persistent Memory

Hermes remembers who you are, your preferences, environment details, and lessons learned across sessions. Correct it once β€” never again.

πŸ“š Self-Improving Skills

Hermes saves reusable procedures as skills that load automatically in future sessions. It literally learns from experience.

🌐 Multi-Platform Gateway

The same agent runs on Telegram, Discord, Slack, WhatsApp, iMessage, Email and a dozen more platforms β€” full tool access everywhere.

πŸ”„ Provider-Agnostic

Works with OpenRouter, Anthropic, OpenAI, Google, DeepSeek, xAI, local models and 20+ others. Swap models mid-workflow.

⏰ Cron & Background Jobs

Schedule recurring tasks, run long missions in the background, chain jobs together β€” durable and retry-safe.

πŸ‘₯ Multi-Agent Delegation

Spawn subagents or independent Hermes processes for parallel workstreams, each with isolated context.

01 Essential slash commands

Type these inside any session:

/status          # session, model, token & context info
/model [name]    # switch model (session-scoped)
/compress        # compress context when it gets long
/skills          # search & install skills interactively
/tools           # enable/disable toolsets
/cron            # manage scheduled tasks
/memory          # review pending memory writes
/new             # fresh session (clears context = saves tokens!)

02 Key CLI commands

hermes                    # interactive chat (default surface)
hermes setup              # guided setup wizard
hermes doctor             # health check dependencies & config
hermes tools list         # enumerate toolsets per platform
hermes skills browse      # explore the skills hub catalog
hermes mcp add NAME ...   # connect an MCP server
hermes desktop            # native desktop app
hermes dashboard          # web admin panel + embedded chat

03 Background systems worth knowing

SystemWhat it does
delegate_taskSpawn subagents with isolated context for parallel subtasks β€” single, batch, or background mode.
cronjobDurable scheduler: "30m", "every monday 9am", cron expressions, ISO timestamps. Jobs can load skills, override models, and deliver to any platform.
CuratorMaintains agent-created skills automatically β€” tracks usage, archives stale ones (never deletes), keeps backups.
KanbanMulti-profile work-queue board for coordinating multiple agents on shared tasks.
Pro tip: Use /goal to set a standing objective that persists across turns, and /steer to inject course corrections mid-task without waiting for the current work to finish.

MCP: Model Context Protocol

Plug external capabilities straight into every conversation.

? What is MCP and why it matters

MCP is an open standard that lets AI agents connect to external tools and data sources through a universal protocol β€” like USB-C for AI capabilities. Instead of writing custom integrations for every service, you install a server once and its tools appear natively in the agent.

1 Install the prerequisites

pip install mcp          # MCP SDK (optional dep β€” without it MCP is disabled)
# Node.js needed for `npx` servers Β· uv needed for `uvx` (Python) servers

2 Find MCP servers

3 Connect a server

Via the CLI (recommended):

hermes mcp add time --command uvx mcp-server-time
hermes mcp add github --command npx -y @modelcontextprotocol/server-github
hermes mcp add myapi --url https://mcp.example.com/mcp
hermes mcp list                # see what's configured
hermes mcp test github         # verify connectivity

Or edit ~/.hermes/config.yaml directly under the mcp_servers key:

mcp_servers:
  time:
    command: "uvx"
    args: ["mcp-server-time"]

  github:
    command: "npx"
    args: ["-y", "@modelcontextprotocol/server-github"]
    env:
      GITHUB_PERSONAL_ACCESS_TOKEN: "ghp_..."   # only passed to THIS server

  remote_api:
    url: "https://mcp.example.com/mcp"
    headers:
      Authorization: "Bearer sk-..."
    timeout: 180

Restart Hermes β€” it connects, discovers tools, and registers them as mcp_{server}_{tool} (e.g. mcp_time_get_current_time). Then just ask naturally: "what time is it in Tokyo?"

Config options reference
  • command / args β€” executable + arguments (stdio transport)
  • env β€” extra env vars passed to the subprocess (secrets go here, explicitly)
  • url / headers β€” remote HTTP transport (use one transport or the other, not both)
  • timeout β€” per-tool-call timeout (default 120s) Β· connect_timeout β€” startup timeout (default 60s)

Toggle which tools from a server are exposed with hermes mcp configure NAME. Reload live with /reload-mcp.

Troubleshooting quick hits
  • "MCP SDK not available" β†’ pip install mcp
  • "No MCP servers configured" β†’ check the key is exactly mcp_servers in config.yaml
  • "Failed to connect" β†’ command not on PATH? package needs -y? bump connect_timeout?
  • Tools not appearing β†’ look for the mcp_{server}_{tool} prefix; check YAML indentation
Security note: Hermes filters environment variables for stdio servers β€” only safe basics like PATH/HOME pass through. Secrets reach a server only if you explicitly put them in its env block. Disable sampling for untrusted servers with sampling: {enabled: false}.

Skills Available

Skills are Hermes' procedural memory β€” reusable playbooks loaded on demand. Search or filter the library below.

Create & Add New Skills

Three ways to grow your skill library β€” install from the hub, learn from experience, or author by hand.

01 Install from the hub

hermes skills browse           # interactive catalog
hermes skills search "pdf"     # find by keyword
hermes skills install ID       # by hub ID...
hermes skills install https://example.com/SKILL.md   # ...or direct URL
hermes skills tap add REPO     # add a GitHub repo as a skill source

02 Let Hermes learn it itself

The most natural way β€” just ask in conversation:

03 Author one by hand

Skills live in ~/.hermes/skills/<name>/SKILL.md (or a category subdirectory). Structure:

---
name: my-skill
description: "Use when X happens. Does Y."   # ← trigger lives here; first ~57 chars show in the index
version: 1.0.0
---

# My Skill

## When to Use
Clear trigger conditions.

## Steps
1. Numbered steps with exact commands.

## Pitfalls
Known gotchas and how to avoid them.
Description is everything. The agent decides whether to load your skill based purely on that line β€” make it self-contained and trigger-focused: "Use when <trigger>. <one-line behavior>."

Then share it: hermes skills publish PATH publishes to the hub. Keep supporting files under references/, templates/, or scripts/ within the skill directory.

04 Managing skills over time

Tip: after finishing a difficult multi-step task, just tell Hermes "save this as a skill" β€” it writes a SKILL.md capturing the workflow, pitfalls included, so future sessions get it right the first time.

Token Efficiency

Context window is money and attention. Ranked tactics by impact.

β˜… Highest impact habits

$ Check your burn rate

/usage shows token usage and rate limits for the session. /insights 7 gives usage analytics over days. /status shows current context fill level β€” compress before you hit the ceiling.

Golden rule: never change past context, toolsets, or the system prompt mid-conversation β€” this breaks prompt caching and re-bills your entire history at full price. Enable/disable tools between sessions (/reset applies tool changes), not mid-chat.