Agentic AI · Self-hosted
Hermes
Self-Hosted Agentic AI Assistant with Autonomous Tool Calling & Headless Browser Automation
- Execution Speed
- <1.2s
- Tool Accuracy
- 98.4%
- Daemon Uptime
- 99.9%
How a request runs
The tool-calling loop
- 1
Natural Prompt Input
User or scheduled cron triggers task via socket or CLI interface.
- 2
LLM Intent & Tool Parser
Hermes reasons over schema registry and chooses exact deterministic tool calls.
- 3
Sandboxed Tool Execution
Executes Python scripts, REST APIs, or Headless Browser sessions in isolated worker.
- 4
Telemetry & Verification
Validates output format, captures telemetry, and returns structured result to user.
The problem
Too many tabs, too many terminals
Enterprise engineers frequently switch context between CLI terminals, web dashboards, and documentation. Traditional chatbots only return plain text and cannot interact with host environments or execute real multi-step workflows safely.
In a busy MSP, engineers run hundreds of small checks a day: API tokens, backup snapshots, CVE advisories, device registration across tenants. Every switch between a portal and a terminal adds delay and room for mistakes.
The build
A sandboxed agent that uses real tools
Architected a modular agent runtime featuring deterministic JSON schema tool-calling, headless Chromium browser control, persistent memory queues, and an encrypted credential manager for API keys.
Hermes runs as an isolated systemd daemon. It listens on local UNIX sockets and REST endpoints and turns questions into strict Pydantic schemas. Each tool call runs in a worker process with tight permissions, and the output comes back as a short, useful summary.
Results
What changed
- Reduced routine infrastructure lookup times by ~60% through automated query-and-fetch commands.
- Implemented zero-trust execution sandbox preventing unauthorized write operations on host systems.
- Engineered autonomous multi-step reasoning loops with self-correcting error recovery.