Hermes Agent - Core Skill Setup Guide¶
Source: nousresearch/hermes-agent (Official)
Skill: hermes-agent · Installs: 400+ · Category: Core / Agent Framework
Platform: Linux, macOS, Windows
The official Hermes Agent core skill from Nous Research. Bundles the essential capabilities every Hermes agent needs: CLI invocation patterns, subagent delegation (delegate_task), persistent memory (FTS5 + LLM summaries), self-improving skills that auto-create from completed tasks, MCP bidirectional integration, browser automation, code execution, and web research.
Installation¶
# Install via skills.sh
npx skills add nousresearch/hermes-agent --skill hermes-agent -g -y
# Or clone the full repo
git clone https://github.com/NousResearch/hermes-agent.git
# Skill is at: hermes-agent/skills/hermes-agent/SKILL.md
Prerequisites¶
| Requirement | Details |
|---|---|
| Hermes Agent | v0.20.0+ |
| Shell Access | hermes CLI must be in PATH |
| API Keys | At least one LLM provider configured in config.yaml |
Core Capabilities¶
1. CLI Invocation¶
Run Hermes from any agent or script:
# One-shot task execution
hermes run "Analyze this CSV file and generate a summary report"
# With specific model
hermes run --model claude-sonnet-4-6 "Review this PR diff"
# With skill pre-loading
hermes run --skill corpusiq-execution-discipline "Audit system health"
2. Subagent Delegation¶
Isolate complex subtasks into independent contexts:
# Delegate research task
hermes delegate "Research top 5 MCP server implementations and compare features"
# Parallel delegation (up to 3 concurrent)
hermes delegate --parallel \
"Check X mentions" \
"Scan Reddit for CorpusIQ threads" \
"Review HN front page for relevant discussions"
Subagents get their own conversation, terminal session, and toolset. Results flow back as summaries - intermediate tool output never floods context.
3. Persistent Memory¶
Two-tier memory architecture:
- FTS5 full-text search - fast keyword/boolean retrieval across all past sessions
- LLM summaries - semantic compression of key facts, preferences, and patterns
# Search past sessions
hermes session search "MCP server deployment"
# Save durable facts
hermes memory add "Project uses PostgreSQL 17 with PGLite for dev"
4. Self-Improving Skills¶
Agents learn from completed tasks:
- After complex tasks (5+ tool calls), the agent proposes a new skill
- Skills encode the workflow, pitfalls, and verification steps
- Next time the task type appears, the agent loads the skill automatically
- Quality improves over time without manual curation
# List available skills
hermes skills list
# Create a new skill from a completed task
hermes skills create --from-session latest
5. MCP Integration¶
Bidirectional Model Context Protocol:
- Client: Connect to external MCP servers (53+ tools registered)
- Server: Expose Hermes capabilities as MCP tools for other agents
# List connected MCP servers
hermes mcp list
# Test a connection
hermes mcp test <server-name>
6. Browser Automation¶
Built-in browser tools for web interaction:
- Navigate, snapshot, click, type, scroll, vision analysis
- Console inspection for JS errors
- Screenshot capture with AI annotation
Common Patterns¶
Pattern: Research → Synthesize → Report¶
# 1. Delegate research (runs in background)
hermes delegate "Research competitor pricing for AI data connectors"
# 2. While research runs, prepare report template
hermes run "Create a competitive analysis report template"
# 3. When research completes, synthesize
hermes run "Merge research findings into the report template"
Pattern: Code Review Pipeline¶
# 1. Delegate deep code review
hermes delegate --skill github-code-review "Review PR #42 in corpusiq-docs"
# 2. Parallel: check for security issues
hermes delegate --skill security-audit "Audit PR #42 for vulnerabilities"
# 3. Merge findings
hermes run "Combine code review and security audit into unified PR feedback"
Pattern: Daily Operations¶
# Morning health check
hermes run --skill corpusiq-system-auditor "Run system health check"
# Midday social sweep
hermes run --skill corpusiq-organic-discovery "Check all platforms for mentions"
# Evening report
hermes run --skill corpusiq-daily-html-reporting "Generate daily report"
Cost Awareness¶
| Task Type | Estimated Cost |
|---|---|
| Simple Q&A (single turn) | $0.001-0.01 |
| Research task (multi-turn) | $0.02-0.10 |
| Complex task with subagents | $0.05-0.50 |
| Full daily operations sweep | $0.10-1.00 |
Costs vary by model selection. Sonnet is the cost-efficient default. Opus reserved for deep research and complex architecture.
Related Skills¶
- Hermes Agent Skill Authoring - Writing SKILL.md files
- dandacompany/hermes - Self-hosting and operations guide
- wihy/hermes-agent-skill - Portable CLI wrapper v2.0
- Hermes Documentation - Official docs
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.