Tech • IA • Crypto
I tested Hermès 2.0 with Claude Code: The result is crazy!
7/10TL;DR
Hermes 2.0 introduces advanced AI agent orchestration, but its real power depends on deep configuration and integration with external models like Claude and MiniMax, rather than out-of-the-box use.
KEY POINTS
Hermes 2.0 Positions Itself as an AI Agent Hub
Hermes 2.0 is presented as a high-performance AI system capable of orchestrating multiple language models, tools, and workflows within a single interface. Rather than being a standalone intelligence, it acts as a wrapper that connects models, memory, tools, and automation logic into a unified architecture. This positioning reflects a broader shift toward modular AI systems rather than monolithic models.
Critical Limitation: No Native Browser Control
A key weakness of Hermes is its lack of full browser control, preventing it from handling tasks such as navigating websites or extracting data from images. This limitation becomes significant in real-world use cases like analyzing Energy Performance Certificate (EPC) images. To overcome this, integration with Claude, which includes a Chromium-based browsing system, enables full web interaction and visual processing.
Claude Integration Expands Capabilities
Hermes now allows direct terminal-based access to Claude 4.8 Code, enabling users to delegate complex browsing and execution tasks. Claude can autonomously navigate websites, extract structured data, and execute workflows without user intervention. This hybrid setup effectively combines Hermes’ orchestration with Claude’s execution layer.
Configuration Is Essential, Not Optional
Despite claims of autonomy, Hermes does not function effectively without manual configuration. Most users fail by simply installing the system and connecting a model. The core issue lies in the default absence of a system prompt file (Sys.md), which must be manually created and populated to define behavior, memory usage, and tool access.
System Prompt Acts as the “Brain”
The Sys.md file determines how Hermes operates, including where memory is stored, which tools are available, and how decisions are made. Without it, the system lacks operational logic. Proper configuration transforms Hermes from an empty shell into a structured agent capable of executing tasks reliably.
MCP Tools Enable Real-World Interaction
Hermes relies on Model Context Protocol (MCP) connections to interact with external services like Airtable and Notion. These integrations can be configured via API keys or OAuth, with OAuth offering simpler setup. Once connected, Hermes can automate workflows such as CRM updates or knowledge base management.
AI Agents Require Full Architectures
Creating effective AI agents involves more than prompting. Functional systems require structured architectures using YAML configurations, scripts, triggers, and validation logic. Simple instructions like “act as an expert” are insufficient, particularly for complex tasks involving multiple data types or tools.
Automation Use Case Demonstrates Practical Value
A typical workflow includes extracting emails tagged “partnership” from Gmail, parsing relevant data, and sending it to Airtable with status tracking. The system can also generate and send replies automatically using predefined templates. These processes can run on schedules via cron-like automation.
Multi-Model Strategy Reduces Costs
Hermes supports multiple providers simultaneously, including MiniMax, Claude, ChatGPT, and Ollama. Cost differences are significant: MiniMax can be up to 10 times cheaper than Claude and 5 times cheaper than OpenAI, while offering comparable performance due to training on Claude-based systems.
Privacy vs Performance Trade-offs
Users concerned about data privacy can opt for Ollama, which runs models without training on user data but incurs higher compute costs. Alternatively, cloud-based providers offer scalability and lower pricing at the expense of data exposure.
Local Deployment Enables Free Usage
Hermes can be used entirely free by running local models via LM Studio. Models such as Gemma 4 can operate on consumer hardware, with requirements ranging from 3B to 12B parameters depending on available VRAM. This setup allows offline processing and full control over data.
Parallel Agents and Real-Time Monitoring
Hermes supports spawning multiple agents simultaneously, enabling parallel task execution. A new interface allows users to monitor these agents in real time through floating windows, improving transparency and multitasking capabilities.
CONCLUSION
Hermes 2.0 highlights the evolution of AI toward orchestrated agent systems, where effectiveness depends less on raw model power and more on structured configuration, tool integration, and multi-model collaboration.