Autonomous DevOps Agent

Point it at any project folder. It writes your Docker / GitHub Actions / Kubernetes files, pushes to GitHub, watches the pipeline, and when something breaks it finds the cause and fixes it — asking for your approval before every step.

devops  (run inside your project folder)
   │
   ├── 1. Scaffold     → Dockerfile, docker-compose, GitHub Actions, k8s manifests
   ├── 2. CI/CD        → pushes to GitHub, watches the Actions run
   ├── 3. Monitoring   → reads the CI logs, finds what failed
   ├── 4. Knowledge    → investigates the cause (knowledge base + web search)
   └── 5. Self-Healing → edits the broken file, shows you the diff, next run verifies it

Works with Groq, Gemini, OpenAI, Claude (cloud) or Ollama (local). Tested on Windows.

⚡ Quick start (5 minutes)

pip install autonomous-devops-agent
cd path/to/your/project        # a folder with your code (and a requirements.txt for Python)
devops --doctor                # checks everything; fix any [FAIL]
devops                         # start

Before the first run you need two keys in a .env file inside your project folder (see Keys) and two secrets in your GitHub repo (see GitHub setup).

Windows tip: if you installed inside a virtual environment, you must activate it in every new terminal before devops works: C:\path\to\.venv\Scripts\activate — otherwise you get "'devops' is not recognized".

📦 Install

Needs Python 3.10+ and git. About 450 MB. 4 GB RAM is enough with a cloud model.

Recommended (isolated):

python -m venv .venv
.venv\Scripts\activate          # Windows
source .venv/bin/activate       # macOS / Linux
pip install autonomous-devops-agent

Or the one-command installer from this repo (creates its own environment and adds devops to your PATH):

./install.sh          # macOS / Linux
.\install.ps1         # Windows PowerShell

The first run downloads a small search model (~90 MB) once. Everything else runs inside the package — no Docker or database server needed.

🔑 Keys

Create a file named .env in your project folder (or once in ~/.devops_agent/.env to use it for all projects):

# Pick at least ONE model provider
GROQ_API_KEY=gsk_...          # free: https://console.groq.com → API Keys
GEMINI_API_KEY=...            # free: https://aistudio.google.com/apikey
OPENAI_API_KEY=sk-...         # https://platform.openai.com/api-keys
ANTHROPIC_API_KEY=sk-ant-...  # https://console.anthropic.com

# Required for pushing to GitHub
GITHUB_TOKEN=ghp_...
GITHUB_REPO=https://github.com/you/your-project.git   # optional — otherwise you're asked once

GitHub token: github.com → Settings → Developer settings → Personal access tokens → Tokens (classic) → Generate → tick repo and workflow.

🔒 Your .env never leaves your computer. Before every push the agent adds .env to your project's .gitignore, and it never stores your token in .git/config. Never paste your .env into chats, issues or screenshots.

Ollama (optional, free, local): install from https://ollama.com, then ollama pull llama3.2:3b, and pick [1] Ollama in the menu. Needs 8 GB+ RAM.

🐙 GitHub setup (once per repo)

1. Create a repo for your project on github.com (empty: no README, no .gitignore, no license). You'll paste its URL when devops asks for it.

The generated pipeline builds your Docker image and pushes it to Docker Hub, so GitHub also needs your Docker Hub login:

2. Create a free account at https://hub.docker.com → Account settings → Personal access tokens → create one.
3. In that GitHub repo: Settings → Secrets and variables → Actions → New repository secret, add:
   • DOCKER_USERNAME = your Docker Hub username
   • DOCKER_PASSWORD = the token from step 1

Without these, the first pipeline fails at docker login — the agent will tell you, but it can't fix secrets for you. After you add them, choose [1] Run pipeline again: since no file changed, the agent asks GitHub to re-run the workflow for you.

▶️ What happens when you run devops

What the agent does to your repo: it commits and pushes to main. If GitHub has newer commits (a teammate, or a repo created with a README), it first puts your commits on top of theirs — nothing on GitHub is ever deleted, and it never force-pushes. Only if the same lines were changed on both sides does it stop and ask you to resolve the conflict once. Fixes are backed up in .self_healing_backups/.

Automatic re-runs: after a fix is applied, the agent pushes it and re-runs CI/CD by itself — incident after incident — until the pipeline is green. It stops and shows the menu when a problem needs you (e.g. Docker Hub secrets, an expired token), when you decline a fix, or when a fix changes nothing new (the same error keeps coming back). You still approve each investigation and fix; Ctrl+C stops the loop. Optional hard limit: DEVOPS_MAX_AUTO_RUNS=10 in .env. The repo URL is remembered, so you only type it once.

Good models to start with:

⚠️ Good to know

• One error per run. GitHub stops at the first failing step, so if your project has 3 bugs the agent fixes them over 3 runs — automatically, one after another. Bugs in the same file are usually fixed together.
• What the pipeline checks: Python syntax (compileall) and the Docker build (including pip install). It does not start your app, so bugs that only appear at runtime aren't detected yet.
• Free-tier limits: 429 (rate limit) and Gemini 503 (busy) are retried automatically; Gemini falls back to a lighter model if busy.

🛠️ Troubleshooting

🗂️ Where things are stored

Email approvals & alerts (optional)

The alert agent sends email notifications and handles approvals via clickable email links. Approvals appear on both your terminal (CLI) and in email simultaneously — first response wins.

Two routing lanes:

• Normal (LOW / MEDIUM severity) → email goes to ALERT_ENGINEER_EMAIL only
• Emergency (HIGH / CRITICAL) → email goes to the full team (ALERT_TEAM_EMAILS) simultaneously

Step 1 — Configure Gmail

1. Google Account → Security → enable 2-Step Verification
2. Go to App Passwords → generate one for "Mail" → copy the 16-character password

Add to .env:

ALERT_SMTP_HOST=smtp.gmail.com
ALERT_SMTP_PORT=587
ALERT_SMTP_USERNAME=you@gmail.com
ALERT_SMTP_PASSWORD=xxxx xxxx xxxx xxxx
ALERT_FROM_ADDRESS=DevOps Agent <you@gmail.com>
ALERT_ENGINEER_EMAIL=lead@company.com
ALERT_TEAM_EMAILS=alice@company.com,bob@company.com,carol@company.com

ALERT_TEAM_EMAILS is a comma-separated list. Leave it blank to disable team broadcasts.

Step 2 — Install Cloudflare Tunnel (Windows)

The approval server needs a public URL so email links work from any device. We use cloudflared (free, no account required).

Install (one-time):

1. Download the Windows binary:

   https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-windows-amd64.exe
   

2. Rename to cloudflared.exe
3. Move it to C:\Windows\System32\ so it's available from any terminal

Verify it works:

cloudflared --version

You should see output like cloudflared version 2024.x.x.

Linux / macOS:

# macOS
brew install cloudflared

# Linux (Debian/Ubuntu)
wget https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64.deb
sudo dpkg -i cloudflared-linux-amd64.deb

Step 3 — Run

No separate server needed. The approval server starts automatically when you run devops. At startup you'll see:

  [Email] approvals/alerts → lead@company.com  |  emergency team → 3 address(es)
  [ApprovalServer] 🌐 Public URL: https://xxxx.trycloudflare.com

When an approval gate is reached, an email is sent to the engineer with ✅ Approve and ❌ Deny buttons. You can also type yes or no directly in the terminal — first response wins.

Dashboard status

When email is configured, the dashboard shows:

● alerting_agent           IDLE (email)

When email is not configured:

○ alerting_agent           —

No cloudflared?

If cloudflared is not installed, the approval server falls back to localhost only — email links will only work if you open them on the same machine running the agent. CLI approval always works regardless.

🐳 Docker Compose (experimental)

Runs everything (agent + Ollama + Qdrant) in containers — only Docker needed. Not yet tested end-to-end; use the pip install above if you're unsure.

cp .env.template .env            # fill in your keys
docker compose up -d             # starts Qdrant + Ollama, pulls llama3.2:3b once
PROJECT_DIR=/path/to/your/project docker compose run --rm devops

Windows PowerShell: $env:PROJECT_DIR="C:\path\to\project"; docker compose run --rm devops

👩‍💻 Development

git clone https://github.com/fatmaabdulfattah/autonomous-devops-agent.git
cd autonomous-devops-agent
pip install -e ".[dev]"          # editable install
pytest tests/test_packaging.py

Upgrade: pip install -U autonomous-devops-agent

autonomous-devops-agent/
├── pyproject.toml              ← package metadata, dependencies, `devops` command
├── install.sh / install.ps1
├── Dockerfile / docker-compose.yml
└── src/auto_devops_agent/
    ├── cli.py                  ← `devops` entry point (--version, --doctor)
    ├── devops_agent/           ← CLI app (pipeline, chat agent)
    ├── core/                   ← orchestrator, event bus, approvals, project DB
    ├── agents/                 ← scaffold · cicd · monitoring · knowledge · self_healing
    └── providers/              ← LLM providers + GitHub provider

License

MIT