Agent Framework Capability Matrix: Hermes → OpenClaw → Claude Code Full Capability Overview

Capability Matrix Overview

Hermes Agent               OpenClaw                     Claude Code
           (Fallback/Background Layer) (Primary Interaction + Orchestration + Deployment Layer) (Multi-Instance Code Execution Layer)
           Model: Qwen 3.6-Plus       Model: DS V4 Pro/Qwen/Local   Multiple instances in parallel
           ModelStudio Coding Plan    oMLX: Qwen3 Coder Next       DS V4 Pro/Qwen/Local
──────────────────────────────────────────────────────────────────────────────
🔍 Search  Tavily ✅                  Tavily ✅                    Tavily ✅
  API      Brave ✅                   Brave ✅                     Brave ✅
           Web Fetch ✅               Web Fetch ✅                 Web Fetch ✅
           For background collection  Real-time search during tasks  Used by OC orchestration
──────────────────────────────────────────────────────────────────────────────
🧠 LLM     Qwen 3.6-Plus              DeepSeek V4 Pro ✅           DeepSeek V4 Pro ✅
  API      (ModelStudio only)         Qwen 3.6-Plus ✅             Qwen 3.6-Plus ✅
                                      oMLX Qwen3 Coder Next ✅     Local Qwen3 Coder ✅
                                      Primary conversation engine  Three instances in parallel
                                      OC selects models for orchestration itself
──────────────────────────────────────────────────────────────────────────────
💰 Finance ─                          In planning ✅               ─
  API                                 HK stocks/financial reports/valuation data
──────────────────────────────────────────────────────────────────────────────
🔌 MCP     ❌ Not involved             ✅ MCP Client                Each instance has its own MCP
  Protocol                            Orchestrates CC's MCP Server  filesystem/github
                                      Bridges for CC               Independent tool set each
──────────────────────────────────────────────────────────────────────────────
🎯 Skills  ❌ No slash commands        ✅ Full slash command system  Slash Commands
  System   Cannot follow instructions precisely  /sdd /finance /reip  /compact /review
           Always "optimizes" and simplifies on its own  /cc /cc-qwen /vercel  Each instance independent
           Autonomous background skill learning  Skills management hub  CLAUDE.md
──────────────────────────────────────────────────────────────────────────────
👤 Role    Fallback: only used when OC can't handle it  Primary entry point (90%+)  Code writing + review
           Background information gathering  Task understanding → decomposition  Multi-instance division of labor:
           System health checks               Resource orchestration + Skills management  · DS V4 Pro: complex logic
           Autonomous skill learning          🚀 Deploy Vercel               · Qwen 3.6: rapid iteration
           Barely a backup                    Sole entry point for slash commands  · Local Coder: private code
──────────────────────────────────────────────────────────────────────────────
📡 Communication  Internal logs/alerts  Feishu / Lark ✅             ─
                                      WhatsApp ✅                   (all through OC)
                                      WeChat ✅
                                      Your only conversation window

Detailed Explanation of Each Layer

Hermes Agent: Backup/Background Layer

Role: Not the one you use daily. It is a background worker plus last-resort backup.

Daily information gathering: Uses Tavily/Brave to gather AI news in the background for you to review
System health checks: Periodically checks whether all components of the Agent framework are running normally
Autonomous skill learning: Absorbs new knowledge and self-updates the skill library
Backup role: Only tasks that OpenClaw cannot handle are passed to it to try

⚠️ Known issues:

No slash commands: Can only communicate via natural language; cannot precisely trigger specific workflows. You are used to slash commands such as /sdd and /vercel, but Hermes does not support them
Unauthorized "optimization": When you tell it to do A and it thinks B is better, it switches to B on its own instead of following your instructions
Over-simplification: Complex instructions are arbitrarily reduced, and the output is worse than expected

Conclusion: Just let it handle background gathering and system checks. Rely on OpenClaw for daily interactions.

OpenClaw: Primary Interaction + Orchestration + Deployment Layer

Role: The one you talk to 90% of the time. The system's central dispatch hub.

Core capabilities:

Capability	Description
Slash command system	`/sdd` `/finance` `/reip` `/cc` `/cc-qwen` `/vercel` are all here
Task orchestration	Receives your instructions → breaks them into subtasks → assigns them to Claude Code
Skills management	Skills such as `claude-code-bridge`, enabling OC to orchestrate CC
Deploy to Vercel	`npx vercel --prod` is executed directly by OC
Multi-model orchestration	Automatically selects DeepSeek V4 Pro / Qwen 3.6 / local Coder based on the task
Multi-channel communication	Feishu/Lark, WhatsApp, and WeChat are all integrated

Large model backends:

DeepSeek V4 Pro (complex reasoning)
Qwen 3.6-Plus via ModelStudio (rapid iteration + Chinese content)
oMLX local Qwen3 Coder Next (privacy-sensitive tasks)

Claude Code: Multi-Instance Code Execution Layer

Role: OpenClaw's "hands". OpenClaw thinks, CC does.

Three parallel instances:

Instance	Model	Purpose
CC DeepSeek	DeepSeek V4 Pro	Complex logic, large-scale refactoring, deep analysis
CC Qwen	Qwen 3.6-Plus	Rapid iteration, UI adjustments, Chinese content
CC Local	oMLX Qwen3 Coder Next	Sensitive code, privacy-focused projects

Responsibilities:

Code writing, review, and refactoring
Shell command execution
Git operations (commit, branch, diff)
Read and edit codebases

A Typical Day

08:00  Hermes quietly gathers information and checks the system in the background (you mostly ignore it)
09:00  You on Feishu → OpenClaw: "/cc review yesterday's commit"
        OC → CC Bridge → CC (DeepSeek V4 Pro) → review report
09:15  You → OpenClaw: "/cc-qwen quickly tweak this UI color"
        OC → CC (Qwen 3.6-Plus) → changes it instantly
10:00  You → OpenClaw: "/vercel"
        OC deploys directly, done in seconds

If OC gets stuck:
  You → Hermes: "Help me try..."
  Hermes: starts unilaterally "optimizing" your instructions 😤
  You: Never mind, wait for OC to recover