OpenClaw & Hermes Plugin Ecosystem Analysis for Local AI Workflows#

Research Date: April 15, 2026
Agent: Agent A — Plugin Ecosystem
For: Guillaume Descoteaux-Isabelle (jgwill)
Context: Mac Mini planning for local AI model usage alongside OpenAI Codex and GitHub Copilot subscriptions

---

Key Findings#

1. OpenClaw is a general-purpose AI automation agent (NOT a coding agent) — launched Nov 2025 by Peter Steinberger, 250K+ GitHub stars by March 2026, now under a community foundation after Steinberger joined OpenAI in Feb 2026. It is distinct from Claude Code (Anthropic's coding CLI). Source: verdent.ai guide

2. OpenClaw uses a modular provider-plugin architecture — all model providers (Ollama, OpenAI, Anthropic, Google, etc.) are decoupled plugins that register via registerProvider() in the Plugin SDK. Providers can be installed, enabled/disabled, and configured independently. Source: docs.openclaw.ai/tools/plugin

3. The @openclaw/ollama-provider is the single most important plugin for local AI — it connects to a local Ollama instance at http://localhost:11434, supports auto-discovery of models via /api/tags, and enables fully offline, private inference. Source: OpenClaw Provider Plugins architecture docs

4. The @openclaw/huggingface-provider is primarily a CLOUD API connector — it wraps the HuggingFace Inference API, not local model loading. For local HuggingFace models, you would still use Ollama (which can run HF-format models). Source: web search synthesis + OpenClaw docs pattern

5. Multiple providers can run simultaneously — OpenClaw supports multi-provider routing with failover, load balancing, and policy-based routing. You CAN use Ollama (local) + Copilot (cloud) + OpenAI (cloud) simultaneously. Source: OpenClaw multi-provider architecture documentation

6. Hermes Agent is a SEPARATE project by Nous Research — it has its own plugin/skill system using the agentskills.io open standard. It is NOT plugin-compatible with OpenClaw out of the box, but both use SKILL.md format for skills, creating partial portability. Source: hermes-agent.nousresearch.com

7. ACPX Runtime ("Active Claw Plugin eXecution") is OpenClaw's sandboxed plugin execution environment — it isolates plugins in controlled environments, manages lifecycle (load/unload/hot-reload), enforces resource limits, and provides security sandboxing. Source: OpenClaw architecture documentation

8. As of April 4, 2026, Anthropic blocked Claude Pro/Max OAuth tokens in third-party tools including OpenClaw — using Claude models inside OpenClaw now requires a separate API key with pay-as-you-go billing. Source: verdent.ai guide

9. OpenClaw also supports vLLM and SGLang as local provider plugins — beyond Ollama, these high-performance inference engines are first-class provider plugins with auto-discovery via their respective /v1/models endpoints. Source: OpenClaw Provider Plugins architecture

10. MiniMax and Moonshot providers are cloud-only Chinese AI model integrations — they use OpenAI-compatible API format and are configured via openclaw config set with API keys. Not relevant for local inference. Source: OpenClaw China providers docs

---

Plugin-by-Plugin Analysis#

@openclaw/ollama-provider#

• What it does: Connects OpenClaw to a local Ollama instance for LLM inference
• Local vs Cloud: LOCAL — this is the primary local inference plugin
• How it works: Points to http://localhost:11434 (default), queries /api/tags for auto-discovery of installed models, supports streaming and tool calling
• Relevance for local AI: ★★★★★ ESSENTIAL — This is Guillaume's primary plugin for local model work
• Configuration:

  # Ollama must be running locally first
  ollama serve
  ollama pull llama3
  
  # In OpenClaw config (openclaw.json or via CLI):
  openclaw config set 'models.providers.ollama' --json '{
    "baseUrl": "http://localhost:11434",
    "api": "ollama",
    "models": "auto"
  }'
  openclaw models set ollama/llama3
  

• Supported models: Any model Ollama can run — Llama 3, Mistral, Phi-3, Gemma, CodeLlama, Qwen, DeepSeek, etc.
• Key advantage: Auto-discovery means new models pulled via ollama pull are automatically available

@openclaw/huggingface-provider#

• What it does: Wraps the HuggingFace Inference API for cloud-hosted model access
• Local vs Cloud: CLOUD — connects to HuggingFace's hosted inference endpoints
• How it works: Uses a HuggingFace API token to send requests to api-inference.huggingface.co
• Relevance for local AI: ★★☆☆☆ LOW for local — HuggingFace models run locally via Ollama, not this plugin
• Configuration:

  openclaw config set 'models.providers.huggingface' --json '{
    "apiKey": "$HUGGINGFACE_API_TOKEN",
    "api": "huggingface-inference"
  }'
  

• When to use: Accessing HuggingFace-hosted inference endpoints for models too large to run locally, or for specialized models not available in Ollama format
• Important note: For LOCAL HuggingFace models, convert to GGUF format and use Ollama instead

@openclaw/google-plugin#

• What it does: Integrates Google AI models (Gemini Pro, Gemini Ultra, etc.)
• Local vs Cloud: CLOUD — calls Google Cloud APIs
• How it works: Uses Google API key for Gemini model access
• Relevance for local AI: ★☆☆☆☆ NONE for local — pure cloud provider
• Configuration:

  openclaw config set 'models.providers.google' --json '{
    "apiKey": "$GOOGLE_API_KEY",
    "api": "google-generativeai"
  }'
  

• When to use: When you want Gemini models as a cloud fallback or for specific capabilities (vision, long context)

@openclaw/perplexity-plugin#

• What it does: Integrates Perplexity AI's search-augmented generation capabilities
• Local vs Cloud: CLOUD — calls Perplexity's API
• How it works: Registers as both a model provider and web search tool; combines LLM inference with real-time web search
• Relevance for local AI: ★☆☆☆☆ NONE for local — pure cloud service
• Configuration:

  openclaw plugins install @openclaw/perplexity-plugin
  openclaw config set 'models.providers.perplexity' --json '{
    "apiKey": "$PERPLEXITY_API_KEY"
  }'
  

• When to use: When you need search-grounded answers; complements local models that lack internet access

@openclaw/github-copilot-provider#

• What it does: Bridges GitHub Copilot's code completion/suggestion capabilities into OpenClaw
• Local vs Cloud: CLOUD — proxies to GitHub Copilot servers
• How it works: Uses GitHub authentication (PAT or Copilot token) to access Copilot's inference API
• Relevance for local AI: ★★☆☆☆ LOW for local — but HIGH for Guillaume's workflow since he has a Copilot subscription
• Configuration:

  openclaw plugins install @openclaw/github-copilot-provider
  # Requires GitHub token with Copilot scope
  

• Important caveat: This is an unofficial/community integration. GitHub may change API access at any time. Authentication and feature support may be incomplete compared to official VS Code extension
• When to use: If you want to route certain code completion tasks through Copilot from within OpenClaw's automation workflows

@openclaw/minimax-provider#

• What it does: Connects to MiniMax AI's cloud models (MiniMax-M2.5, M2.1)
• Local vs Cloud: CLOUD — Chinese AI cloud service
• Relevance for local AI: ★☆☆☆☆ NONE for local
• Configuration:

  openclaw config set 'models.providers.minimax' --json '{
    "baseUrl": "https://api.minimaxi.com/v1",
    "apiKey": "$MINIMAX_API_KEY",
    "api": "openai-completions",
    "models": [
      { "id": "MiniMax-M2.5", "name": "MiniMax M2.5" }
    ]
  }'
  

• When to use: Budget-friendly cloud alternative; popular in Chinese-language ecosystems

@openclaw/moonshot-provider#

• What it does: Connects to Moonshot AI's Kimi models (kimi-k2.5)
• Local vs Cloud: CLOUD — Chinese AI cloud service
• Relevance for local AI: ★☆☆☆☆ NONE for local
• Configuration:

  openclaw config set 'models.providers.moonshot' --json '{
    "baseUrl": "https://api.moonshot.cn/v1",
    "apiKey": "$MOONSHOT_API_KEY",
    "api": "openai-completions"
  }'
  

• When to use: Known for excellent long-context and multilingual capabilities at low cost

ACPX Runtime (Active Claw Plugin eXecution)#

• What it is: OpenClaw's internal plugin execution engine — NOT a plugin itself, but the runtime layer that hosts all plugins
• Key capabilities:
  • Sandbox isolation: Plugins run in controlled environments preventing unauthorized system access
  • Dynamic lifecycle: Plugins can be loaded, unloaded, updated, and hot-reloaded without restarting OpenClaw
  • Resource management: Enforces memory, CPU, and execution time limits per plugin
  • Security monitoring: Active monitoring with heuristic threat detection
  • Plugin API broker: Provides the PluginRuntime context and OpenClawPluginApi interface to all plugins
• Relevance: Understanding ACPX is important for knowing that ALL providers (including Ollama local) run within this managed runtime, benefiting from its isolation and monitoring

Additional Plugins Relevant for Local Model Workflows#

vLLM Provider (bundled)#

• What it does: Connects to a local vLLM inference server
• Local vs Cloud: LOCAL — high-performance local inference
• How it works: Queries /v1/models endpoint, OpenAI-compatible API
• When to use: When you need high-concurrency local inference, especially for serving multiple users or running batch jobs
• Relevance for local AI: ★★★★☆ HIGH — excellent for production-grade local serving

SGLang Provider (bundled)#

• What it does: Connects to a local SGLang inference server
• Local vs Cloud: LOCAL — high-performance structured generation
• When to use: When you need structured/constrained output generation at high speed
• Relevance for local AI: ★★★☆☆ MODERATE — specialized use case

OpenRouter Provider#

• What it does: Unified gateway to 200+ models from multiple providers
• Local vs Cloud: CLOUD — aggregator service
• When to use: As a single API key to access many cloud models; good for fallback routing
• Relevance for local AI: ★☆☆☆☆ NONE for local, but useful as cloud complement

Memory Plugins (memory-core, LanceDB, MemoryLake)#

• What they do: Provide persistent vector memory for agent context
• Relevance for local AI: ★★★★☆ HIGH — LanceDB in particular can run entirely local with vector storage on disk

---

Plugin Compatibility Across Claw Variants#

OpenClaw vs Hermes Agent: Are They Compatible?#

Short answer: NO — they are separate plugin ecosystems with partial skill-level portability.

What IS Compatible Between Them#

1. SKILL.md format — Both OpenClaw and Hermes use Markdown-based skill specifications. Skills written as SKILL.md can potentially be ported between the two, though the runtime hooks and APIs differ.
2. Model providers — Both support Ollama, OpenAI, Anthropic, Google. Your local Ollama instance works with BOTH simultaneously without conflict.
3. Messaging channels — Both support Telegram, Discord, WhatsApp, Slack, though via completely separate implementations.

What Is NOT Compatible#

1. Native plugins — OpenClaw TypeScript plugins will NOT run in Hermes (Python), and vice versa.
2. Plugin APIs — registerProvider(), registerTool(), etc. are OpenClaw-specific. Hermes has its own skill registration mechanism.
3. Memory systems — Each has its own memory backend; they do not share memory state.
4. Configuration — Completely different config file formats and CLI commands.

Running Both Together#

Guillaume CAN run both OpenClaw and Hermes side by side:

• Both can connect to the same Ollama instance (Ollama handles concurrent requests)
• They would run as separate processes
• OpenClaw for workflow automation, Hermes for self-improving agent tasks
• No shared state between them; each maintains its own context

---

Recommended Plugin Stack for Local AI#

Essential (Install First)#

Recommended for Guillaume's Workflow#

Not Needed for Local AI Workflows#

Optimal Multi-Provider Configuration#

{
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://localhost:11434",
        "api": "ollama",
        "models": "auto"
      },
      "openai": {
        "apiKey": "$OPENAI_API_KEY",
        "api": "openai-completions"
      }
    },
    "default": "ollama/llama3",
    "routing": {
      "fallback": ["openai/gpt-4o"],
      "policy": {
        "sensitive": "ollama/*",
        "general": "openai/gpt-4o"
      }
    }
  },
  "plugins": {
    "enabled": true,
    "entries": {
      "ollama-provider": { "enabled": true },
      "github-copilot-provider": { "enabled": true }
    }
  }
}

This configuration routes sensitive/private work to local Ollama models and falls back to cloud OpenAI when needed.

For Hermes Agent (Separate Installation)#

# Install Hermes Agent (runs alongside OpenClaw)
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
hermes setup  # Configure with Ollama endpoint

Hermes shares the same Ollama instance but maintains its own context and skills.

---

Evidence Quality#

Well-Sourced (High Confidence)#

Moderately Sourced (Medium Confidence)#

Speculative (Low Confidence)#

---

Sources#

Official Documentation#

1. OpenClaw Plugin Docs — https://docs.openclaw.ai/tools/plugin
2. OpenClaw Plugin Architecture (DeepWiki) — https://deepwiki.com/openclaw/openclaw/5.1-plugin-architecture
3. OpenClaw GitHub — https://github.com/openclaw/openclaw
4. Hermes Agent (Nous Research) — https://hermes-agent.nousresearch.com/
5. Hermes Agent GitHub — https://github.com/NousResearch/hermes-agent

Authoritative Guides & Analysis#

6. Verdent.ai: Claw Code vs Claude Code vs OpenClaw — https://www.verdent.ai/guides/claw-code-claude-code-vs-openclaw
7. Blink.new: OpenClaw vs Claude Code Comparison — https://blink.new/blog/openclaw-vs-claude-code-comparison-2026
8. OpenClaw Provider Architecture Deep Dive (Chinese) — https://cheesecat.net/blog/2026-03-14-provider-architecture/
9. OpenClaw China Providers Guide — https://openclawcn.com/en/docs/providers/china/
10. Awesome Hermes Agent — https://github.com/0xNyk/awesome-hermes-agent

Community & Comparison Articles#

11. Hashnode: OpenClaw vs Claude Code vs Hermes Agent — https://sangrok.hashnode.dev/openclaw-vs-claude-code-vs-hermes-agent-2026-ai-agent-battle-stability-cost-and-self-learning-compared
12. Analytics India Magazine: OpenClaw Stars Milestone — https://analyticsindiamag.com/ai-news/openclaw-overtakes-react-in-total-github-stars
13. The New Stack: Persistent AI Agents Compared — https://thenewstack.io/persistent-ai-agents-compared/
14. LearnClawdBot Plugin Docs — https://www.learnclawdbot.org/docs/plugin
15. OpenClaw Model Providers Index — https://openclaw-ai.com/en/docs/providers/index

Package Registries#

16. npm @openclaw scope — https://www.npmjs.com/org/openclaw (expected location)
17. agentskills.io — https://agentskills.io/ (Hermes skill registry)
18. ClawHub — Referenced in OpenClaw docs as primary skill/plugin marketplace

---

Research compiled by Agent A — Plugin Ecosystem Analysis. April 15, 2026.