YOU OR AN EVENT
Starts work
Chat, editor, scheduled job, webhook, API request, or messaging platform.
OPERATOR'S FIELD GUIDE · JULY 2026
Hermes,
explained.
Hermes is the operating environment around an AI model: it gives that model tools, memory, rules, profiles, schedules, safety controls, and places to work. A model can answer. Hermes can investigate, edit, verify, remember, automate, and hand work off.
agent + toolspersistent memoryprofilesautomationmulti-provider
The mental model
HERMES RUNTIME
Shapes work
Profiles, SOUL/project rules, tools, memory, approvals, context management, skills, and routing.
MODELS + SYSTEMS
Does work
LLM providers plus terminal, files, web, browser, vision, services, and integrations.
Why it is distinctive: Hermes is built to compound. It can preserve useful preferences, encode repeat procedures as skills, isolate different operating roles into profiles, and run durable automation—not merely restart a fresh chatbot every turn.
Choose the right work primitive
| Primitive | Use it when | Do not use it when |
|---|---|---|
| Chat | You are steering interactively and the work belongs in this conversation. | You need a durable/repeating workflow. |
| Delegation | A few short independent tasks need to return to the parent now. | The work must survive restarts or human review. |
| Cron | Work is time-based. Use no-agent scripts for deterministic watchdogs. | An external event should trigger it immediately. |
| Webhook | GitHub, CI, monitoring, or an IoT/service event should start a narrow response. | You only need periodic polling. |
| Kanban | Work crosses agents, needs dependencies/retries/review, survives restarts, or needs a lasting trail. | You merely need one child answer before continuing. |
| Skill / bundle | A procedure keeps recurring and deserves a reusable playbook. | A one-off fact belongs in memory instead. |
Kanban, plainly: delegation is a function call; Kanban is a durable work queue. Named workers can claim tasks, leave comments, block/unblock, retry, hand off to another profile, and preserve an audit trail. Use it for a research → implementation → review pipeline, a fleet of monitored services, or work that may outlive this session.
Feature atlas
IDENTITY + CONTEXT
Profiles, SOUL, project rules
Profiles isolate config, sessions, skills, memory, cron, and gateway state. SOUL.md defines operational stance; AGENTS.md/CLAUDE.md and similar files teach the agent how a particular project works.
Use: coding, research, and ops should not share one undifferentiated brain.
MODELS
Providers, fallbacks, helpers
Set a profile default; override one session/turn; pin cron jobs; route helper functions such as titles, approval, extraction, curator, and vision separately. Fallback providers protect against outages.
Use: quality/cost/reliability differ by task.
TOOLS
Toolsets + approvals
Enable terminal, files, web, browser, vision, skills, memory, delegation, and more per profile. Smart approvals and hard loop stops contain side effects and failure spirals.
Use: grant only the capabilities a role needs.
MEMORY
Built-in memory, sessions, Mem0
Keep durable preferences, topology, recurring workarounds, and decisions. Session search recalls what happened. External semantic memory can add retrieval and extraction.
Avoid: storing routine health-check noise as knowledge.
WEB + MEDIA
Search, extraction, browser, vision, voice
Search finds sources; extraction reads them; browser handles interactive sites; vision understands screenshots; voice supports mobile/gateway interaction.
Use: when the truth is outside the host.
AUTOMATION
Cron, hooks, plugins, webhooks
Cron is clock-driven. Webhooks are event-driven. Hooks react to lifecycle points. Plugins add custom tools and stable integrations without modifying Hermes core.
Use: turn proven recurring work into a system.
INTEGRATION
MCP, proxy, API server, ACP
MCP imports external tools. The proxy provides raw OpenAI-compatible inference to clients. The API server exposes the full tool-using agent. ACP puts Hermes inside compatible editors.
Use: connect one real service at a time, narrowly.
RECOVERY
Checkpoints, backups, logs, insights
Checkpoints snapshot before writes; backups preserve Hermes state; logs reveal failures; insights and usage files show cost, tools, sessions, and model behavior.
Use: builder checkpoints are high-value; cloud tracing can wait.
Settings that matter
model / auxiliary / fallback_providers
Intelligence, cost, latency, helper functions, and resilience. This is where profile defaults and small/default/large tiers belong.
agent / terminal / approvals / security
Turn limits, verification, where commands execute, approval policy, redaction, lazy installs, and loop guardrails.
compression / caching / memory / sessions
Context efficiency and recall. Change these only for observed pain: defaults often protect caching and prevent lossy summary churn.
cron / kanban / curator / hooks / webhooks
Scheduled work, durable workflows, skill maintenance, lifecycle reactions, and event triggers. Pin important automation to a profile/model/toolset/workdir.
mcp_servers / gateway / api_server / proxy
External services and surfaces. Treat each as a trust-boundary decision, not a feature checklist.
checkpoints / backups / insights
Recovery and visibility. Use bounded checkpoints for coding, periodic backups for Hermes state, and native insights before adding observability infrastructure.
Your current glowup
IN PLACE
Separate default, builder, research, and ops-safe profiles; profile-relative small/default/large model tiers; OpenAI primary, Claude fallback, Gemini helper routing; eight isolated ops cron jobs; constrained ops toolset; hard loop stops; builder verification.
NEXT REAL LEVERS
1. One signed webhook for a real alert source. 2. Builder checkpoints for rollback protection. 3. One read-only MCP for GitHub or homelab data. 4. One secure remote channel if you want mobile operations. Do not enable all of them merely because they exist.
A sane operating rule
Enable a bell only when it replaces proven friction. Use webhooks for real events, MCP for a service you repeatedly query, Kanban for durable handoffs, checkpoints before autonomous code edits, and a gateway channel for remote operations. Leave global sandboxing, extra web vendors, cloud tracing, and mixture-of-agents experiments alone until measurement demonstrates the need.