How to Run Hermes Agents Cheaply with DeepSeek

TL;DR

Use Hermes as a persistent coordinator, DeepSeek V4 for inexpensive heavy lifting, and a frontier model such as Claude Opus 4.7 for planning and review. Connecting them through OpenRouter creates a multi-model workflow that can switch tools dynamically, track usage centrally, and run background tasks overnight while reserving the most expensive model for work requiring its highest capability.

Transcript

When you combine Hermes Agent with tools like Deepseek, you unlock capabilities that 99% of people don't even realize exist. And I'm going to show you exactly how to connect Hermes with the world's most powerful models, which means that you can use Hermes from $0. Build as much as you want to with no limitations or rate limits and my new system for... Read More

Key Insights

• Hermes is a persistent agent designed to work across a user's broader life rather than remain confined to one code repository or one session. It learns from assigned tasks, schedules background jobs, and becomes more useful as it develops a deeper understanding of the user.
• Claude Code and Hermes serve different roles in the proposed workflow. Claude Code is described as repository-focused, session-bound, and built around a tight tool loop, while Hermes coordinates persistent tasks, personal context, scheduled work, and connections to other models and tools.
• The multi-brain system assigns different responsibilities to different models. A highly capable frontier model organizes and delegates, DeepSeek V4 handles large volumes of execution, and another capable model critiques the output so that no single model works in isolation.
• DeepSeek V4 is presented as offering roughly 95% of the relevant performance for about 1% of the cost. The practical argument is not that it exceeds Claude Opus 4.7, but that it can complete extensive overnight work without requiring maximum model capability for every step.
• OpenRouter is a unified gateway for accessing and switching among multiple models with one key. Its dashboard centralizes usage tracking, while its routing features can select fast providers, identify providers certified for tool-calling accuracy, choose a model automatically, and provide fallbacks.
• Gemini CLI gives Hermes access to Gemini through a locally installed command-line interface. In the demonstrated task, Hermes uses it to find a YouTube channel and visually analyze movement and other details from the first 10 seconds of a recent video.
• The triad workflow consists of planning, execution, and critique. Claude Opus 4.7 is proposed for rigorous planning, DeepSeek V4 for sustained heavy work, and a highly capable reviewer for severe, critical evaluation before the finished work is shipped.
• OpenRouter supports bringing personal provider keys when a model encounters rate limits. The transcript also describes zero-completion protection, meaning users are not charged for blank or error responses, alongside routing expressions intended to improve speed or tool-calling reliability.
• More videos from this channel:
  • How to build 3D websites with anti gravity and Spline
  • How to Run Claude Code With Free Model APIs
  • How to build gorgeous websites with Kimmi K3 prompts
• More from this channel: Jack Roberts

Questions & Answers

Q: How do Hermes and DeepSeek V4 work together?

Hermes acts as the persistent coordinator that understands the user, schedules background jobs, and selects tools or models for particular tasks. DeepSeek V4 serves as the inexpensive execution engine for large amounts of work. A more capable frontier model can create the plan and later review the result, allowing DeepSeek to continue heavy processing overnight without using the most expensive model throughout.

Q: What is the difference between Hermes and Claude Code?

Claude Code is described as a repository-focused system with a tight tool loop that operates within bounded sessions and is designed for codebases. Hermes is presented as a persistent agent spanning a user's broader life. It learns from each task, schedules background work, builds an understanding of the user, and connects with other models or coding systems when needed.

Q: What is the Hermes triad system?

The triad is a multi-model workflow organized around three stages: plan, execute, and critique. A highly capable model such as Claude Opus 4.7 creates the plan and delegates the work. DeepSeek V4 performs the extensive execution at a lower claimed cost. A strong model then critically reviews the output before it is approved or shipped.

Q: Why use DeepSeek V4 for overnight AI tasks?

DeepSeek V4 is positioned as suitable for overnight tasks because it allegedly provides about 95% of the relevant value for 1% of the cost of the frontier alternative discussed. That cost difference makes sustained background execution more practical. The frontier model can be reserved for planning, delegation, and difficult judgment instead of processing every token of routine work.

Q: How does OpenRouter simplify a multi-model Hermes setup?

OpenRouter gives Hermes a single access point for many models instead of requiring separate keys and usage dashboards for every provider. It enables centralized usage tracking and dynamic model switching. The described features also include automatic model selection, routing toward faster providers, filtering for tool-calling accuracy, fallback providers, personal API key support, and protection from charges for blank responses.

Q: How can Hermes use Gemini CLI for video analysis?

Gemini CLI can be installed on the same computer where Hermes is running. Once available locally, Hermes can be instructed to invoke it for a suitable task. The demonstration asks Gemini CLI to locate Jack Roberts's latest YouTube video and visually analyze the first 10 seconds, including details about movement, taking advantage of Gemini's multimodal analysis capabilities.

Q: What do OpenRouter routing expressions do?

OpenRouter routing expressions modify how a requested model is served. The transcript describes Nitro as routing toward the fastest available provider, Exacto as limiting selection to providers certified for tool-calling accuracy, and OpenRouter Auto as choosing a model for the prompt. These options help Hermes match provider speed, model selection, and tool reliability to the requirements of each task.

Q: How should models be assigned roles in a Hermes workflow?

Models should be assigned according to their particular strengths instead of making one model handle every stage. The proposed setup gives a frontier model the planning and delegation role, sends high-volume execution to DeepSeek V4, and uses a capable model for severe critique. Gemini can separately handle multimodal work, such as visual analysis, through its command-line interface.

Summary & Key Takeaways

• Hermes is presented as a persistent agent that operates across many areas of a user's life, learns from completed tasks, schedules background jobs, and develops a deeper understanding of the user. Unlike repository-focused coding tools, it can coordinate models and local command-line tools for broader, continuing workflows.

• The proposed multi-brain setup assigns each model work that matches its strengths. A frontier model acts as the conductor for planning and delegation, DeepSeek V4 performs extensive background execution at a much lower claimed cost, and another strong model critically reviews the result before anything is shipped.

• OpenRouter provides one access point for multiple models, centralized usage tracking, provider routing, tool-calling filters, fallback options, and support for bringing personal API keys. Hermes can also invoke locally installed tools such as Gemini CLI, allowing it to use Gemini's multimodal capabilities for tasks including visual analysis of YouTube footage.