Mavenix is a self-hosted, agentic AI system specialized in operations, workflow automation, and cross-system execution — built on OpenClaw and connected to any OpenAI-compatible model gateway of your choice.
Unlike general-purpose chatbots, Mavenix is scoped to a single discipline: turning plans into reliable, repeatable, measurable operations. It automates workflows, orchestrates multi-step processes, integrates systems, and monitors execution — reachable from Telegram, Discord, WhatsApp, or Slack.
| Area | What It Does |
|---|---|
| Workflow Automation | Designs and implements automated workflows for repetitive tasks |
| Process Orchestration | Coordinates multi-step operations across systems |
| System Integration | Connects APIs, services, and platforms into unified pipelines |
| Operational Monitoring | Tracks workflow health, detects failures, reports performance |
| Browser & Computer Automation | Executes browser-based and desktop automation tasks |
| Category | Options |
|---|---|
| Messaging | Telegram, Discord, WhatsApp, Slack — configure one or several |
| Model Providers | Any OpenAI-compatible endpoint: OpenAI, Anthropic (via proxy), Google Gemini, Groq, DeepSeek, local models (Ollama, vLLM, llama.cpp), or a routing layer |
| Search | DuckDuckGo (free, no key), Brave Search, Exa, Tavily, SearXNG (self-hosted) |
| Automation | Browser control (Chrome), Computer Use, Google Meet integration |
| Memory | Vector-based persistent memory (LanceDB) with auto-capture and auto-recall |
| Skills | Community skills via skills.sh — see Skills Setup below |
Mavenix doesn't lock you into one paid API. Instead of hardcoding a single provider, point MODEL_GATEWAY_BASE_URL at a routing layer (such as 9Router or similar OpenAI-compatible gateways) that sits between Mavenix and your actual model providers.
This unlocks a few practical things:
- Combo/fallback chains — configure a primary model with automatic fallbacks (e.g., a fast model first, a stronger one if it fails or rate-limits), so the agent keeps working even if one provider hiccups
- Mix free and paid tiers — route through generous free tiers (Google Gemini API's free tier, Groq's free tier, etc.) alongside paid ones, letting the router pick based on availability or cost
- Provider-agnostic — swap providers later without touching Mavenix's code, since it only ever talks to your router's single OpenAI-compatible endpoint
Example: many people configure a free-tier Gemini API key as one node in their router's combo, alongside other free or low-cost providers, to substantially extend how much usage they get before hitting a paywall. This is a router-level decision — Mavenix itself just calls whatever MODEL_GATEWAY_BASE_URL you give it. Check your router's own documentation for how to configure combos and fallback chains; this repository doesn't manage that logic itself.
Mavenix can use community-built Agent Skills — reusable capabilities like SEO audits, PDF processing, data analysis, and more, maintained by an open registry (originally curated by Vercel Labs and community contributors).
Skills are installed via npx from inside the container (or from the host if you mount a shared skills directory — see below):
# Search for a skill
npx skills find "browser testing"
# Install a specific skill
npx skills add browserbase/skills/browser
# List installed skills
npx skills listA reasonable starting set for an operations-focused agent:
# Frontend & Browser Automation
npx skills add vercel-labs/agent-skills/frontend-design
npx skills add browserbase/skills/browser
# Document Processing (Anthropic Skills)
npx skills add anthropics/skills/pdf
npx skills add anthropics/skills/docx
npx skills add anthropics/skills/xlsx
npx skills add anthropics/skills/pptx
# Engineering Practices
npx skills add anthropics/skills/systematic-debugging
npx skills add anthropics/skills/test-driven-development
# Analytics & Workflow
npx skills add community/skills/seo-audit
npx skills add community/skills/web-scraping
npx skills add community/skills/data-analysis
npx skills add community/skills/github-pr-workflow# Execute Batch Install
npx skills add garrytan/gstack --full-depth && \
npx skills add addyosmani/agent-skills && \
npx skills add mattpocock/skills && \
npx skills add supabase/agent-skills && \
npx skills add microsoft/playwright-cli && \
npx skills add tailwindlabs/agent-skills && \
npx skills add docker/agent-skills && \
npx skills add cloudflare/agent-skills && \
npx skills add community/skills/fastapi && \
npx skills add langchain-ai/agent-skills && \
npx skills add Leonxlnx/taste-skill
⚠️ Verify exact package names on skills.sh before installing — the registry is community-maintained and names/availability change. Search first withnpx skills findrather than assuming a name exists.
If you already have skills installed for other tools (Claude, OpenCode, Vercel CLI, etc.) sitting somewhere like C:\Users\<you>\.agents\skills, Mavenix can read them too — without ever mounting or modifying your C: drive.
The approach: copy or symlink the skills folder to D:\ or E:\ first, then mount only that copy into the container as read-only. This keeps your system drive completely untouched while still giving the agent access.
Windows (PowerShell) — one-time copy:
# Copy your existing skills folder to a drive you're comfortable mounting
Copy-Item -Recurse "C:\Users\<you>\.agents\skills" "E:\Mavenix-skills"Then mount that copy (read-only) when running the container:
docker run -d `
--name mavenix `
--restart unless-stopped `
-v ${PWD}:/home/node/.openclaw `
-v E:\Mavenix-skills:/opt/access/skills:ro `
-p 18789:18789 `
mavenix gatewayThe :ro suffix makes the mount read-only — Mavenix can see and use the skills but can't modify or delete anything in that folder. C:\ is never referenced in the mount at all, so there's no path by which the container touches your system drive.
macOS/Linux equivalent:
cp -r ~/.agents/skills /path/on/external-or-secondary-drive/mavenix-skills
docker run -d \
--name mavenix \
--restart unless-stopped \
-v "$(pwd)/data:/home/node/.openclaw" \
-v /path/on/external-or-secondary-drive/mavenix-skills:/opt/access/skills:ro \
-p 18789:18789 \
mavenix gatewayIf you'd rather not maintain a copy, a symlink works the same way — just make sure the symlink target also lives outside C:\.
Installed skills run with the full permissions of the agent — filesystem access, API keys, credentials, shell execution. This is not a sandboxed plugin model. A 2026 security incident on OpenClaw's own skill marketplace (ClawHub) involved hundreds of malicious skills sharing a single command-and-control server, targeting API keys and credentials on installed machines.
Before installing any skill:
- Prefer skills from well-known maintainers (
anthropics/,vercel-labs/, verified orgs) over unknown authors - Read the skill's source/
SKILL.mdbefore installing when possible - Don't install skills you don't understand the purpose of
- Keep the read-only mount pattern above for any pre-existing skill folders — it limits blast radius even if a skill misbehaves
graph TB
User[Telegram / Discord / WhatsApp / Slack] -->|Message| Gateway[Mavenix<br/>Docker Container]
Gateway -->|OpenAI-compatible API| Router[Model Gateway<br/>e.g. 9Router or similar]
Router -->|Route| M1[Gemini API - free tier]
Router -->|Route| M2[Other providers]
Router -->|Route| M3[Local models]
Gateway --> Memory[(Vector Memory<br/>LanceDB)]
Gateway --> Shell[Sandboxed Shell<br/>Execution]
Gateway --> Browser[Browser Automation]
Gateway --> SkillsFolder[(Skills - read-only mount<br/>D:/ or E:/, never C:/)]
Gateway -->|Volume Mount| Workspace[(Persistent Workspace)]
style Gateway fill:#f97316,color:#fff
style Router fill:#6366f1,color:#fff
style Memory fill:#10b981,color:#fff
- Docker installed and running
- A model gateway reachable via an OpenAI-compatible endpoint
- A bot token for at least one messaging platform (Telegram, Discord, WhatsApp, or Slack)
git clone https://github.com/Maventlabs/Mavenix.git
cd Mavenix
cp .env.example .env
nano .env # fill in MODEL_GATEWAY_BASE_URL, MODEL_ID, API key, and your messaging token(s)
docker build -t mavenix .
docker run --rm -it \
-v "$(pwd)/data:/home/node/.openclaw" \
mavenix openclaw onboard
docker run -d \
--name mavenix \
--restart unless-stopped \
-v "$(pwd)/data:/home/node/.openclaw" \
-p 18789:18789 \
mavenix gatewayOn Linux, if your model gateway runs natively on the host, add
--add-host=host.docker.internal:host-gatewayto the onboarding and run commands.
git clone https://github.com/Maventlabs/Mavenix.git
cd Mavenix
Copy-Item .env.example .env
notepad .env
docker build -t mavenix .
docker run --rm -it `
-v ${PWD}:/home/node/.openclaw `
mavenix openclaw onboard
docker run -d `
--name mavenix `
--restart unless-stopped `
-v ${PWD}:/home/node/.openclaw `
-p 18789:18789 `
mavenix gatewayIf your model gateway runs natively on Windows, use
http://host.docker.internal:PORT/v1— notlocalhost— as the base URL.
ssh user@your-vps-ip
git clone https://github.com/Maventlabs/Mavenix.git
cd Mavenix
# ... same steps as Linux deployment aboveMinimum recommended specs: 1 vCPU, 1–2GB RAM. Any Docker-compatible VPS provider works.
docker compose up -dDuring openclaw onboard, choose Manual / Custom when prompted for a model provider, then supply:
- Base URL — your gateway's OpenAI-compatible endpoint
- API Key — your gateway's key
- Model ID — the exact model or combo name your gateway expects (see
MODEL_IDin.env.example— this field is required, not optional)
SOUL.md defines the agent's role and communication style, scoped purely to Operations & Automation with no assumptions about who's using it. To personalize:
- Add a section about yourself so the agent doesn't over-explain things you already know
- Optionally add a Voice & Persona section for a distinct personality
- Fill in your infrastructure details (data folder, model provider, active tools, skills mounted)
- Personal-by-default — one trusted operator boundary
- Bind the gateway to loopback (127.0.0.1) unless you need LAN/remote access
- Never commit
.env,data/, or any credentials (see.gitignore) - Skills run with full agent permissions — see Skills Security Warning above
- Mounted directories (skills or otherwise) should stay off
C:\/system drives — use the read-only pattern shown above - Run periodically:
docker exec mavenix openclaw security audit --deep
.
├── Dockerfile
├── docker-compose.yml
├── .env.example
├── .env # NEVER commit
├── .gitignore
├── SOUL.md
├── data/ # NEVER commit
├── LICENSE
├── CHANGELOG.md
├── CONTRIBUTING.md
└── README.md
Container can't reach your model gateway
Use host.docker.internal instead of localhost on Windows/macOS. On Linux, add --add-host=host.docker.internal:host-gateway.
Onboarding wizard exits unexpectedly
Try ghcr.io/openclaw/openclaw:main instead of :latest.
Port 18789 already in use
Running multiple agents? Give each a distinct port in both the -p mapping and the gateway port during onboarding.
Skill install fails or behaves unexpectedly
Run npx skills list to confirm what's actually installed, and check the skill's SKILL.md for its own requirements — some skills need additional API keys or dependencies not covered by this template.
See CONTRIBUTING.md.
MIT License. OpenClaw itself is a separate project with its own license.
See CHANGELOG.md.