DevOps Agent MCPサーバー

DevOps Agent

An AI DevOps assistant you talk to in plain English. It works on your own AWS, Azure and GCP accounts, Kubernetes clusters, CI/CD systems and on-call tools, with a safety check in front of every risky action.

• Live app: devopsagent.io
• Docs: devopsagent.io/docs (in the app) and the docs/ folder
• Developer API: devopsagent.io/developers

What it does

• Chat with your infrastructure. Ask questions or give instructions ("list my EC2 instances", "scale the api deployment to 4 replicas", "why did last night's deploy fail?"). The agent picks from 749 tools across 18 modules and runs them against your own accounts.
• Workflows. 46 pre-built, step-by-step jobs (launch a web server, audit IAM, troubleshoot pods, self-host an LLM, re-run a failed pipeline, page on-call, and more) across 14 categories. Each opens a short form, then runs in chat.
• Safety. Every tool call is risk-classified. Deletes, terminations, IAM changes and other risky operations need approval before they run (see Approvals). Sandbox mode simulates everything without touching a real account.
• Your data stays yours. Cloud and integration credentials are stored per user, encrypted, and never shared with other users. The app never falls back to its own credentials.

Area	Tools
AWS (EC2, VPC, IAM, RDS, Lambda, ECS, EKS, S3, CloudFormation, Route 53, and more)	352
Kubernetes	76
Git	47
GCP	42
Azure	38
CI/CD (Jenkins, GitHub Actions, GitLab)	36
Terraform	26
Docker	22
GitHub	19
PagerDuty and OpsGenie	16
Ansible	13
Prometheus, Datadog and Grafana	12
SonarQube	12
Nexus	12
Security audit / pentest (off by default)	12
Cost, compliance and change governance	10
Shell commands on the agent host	2
Reachability probes (http_check, tcp_check)	2
Total	749

Using the app

1. Sign up at devopsagent.io and confirm your email.
2. Connect a cloud account in Settings → Cloud Credentials. For AWS, the recommended option is an IAM role (one-click CloudFormation quick-create); access keys also work.
3. Connect any third-party tools you use (GitHub, GitLab, Jenkins, PagerDuty, OpsGenie, Datadog, SonarQube, Nexus) in Settings → Integrations.
4. Chat, or open Workflows and pick one.

Want to try it without a cloud account? Turn on Sandbox mode (Settings → Learning). Tool results are simulated and labelled [Sandbox].

Plans: Free (20 credits a month), Starter ($29, 200 credits), Professional ($99, 1,000 credits) and Business ($299, 3,000 credits). Free and Starter use Kimi K3.1 Pro; Professional and Business use Claude. Extra credit packs are available on the Billing page.

Other ways in

• Claude.ai connector (MCP): add https://devopsagent.io/mcp as a custom connector and sign in with your DevOps Agent account (and 2FA code, if you use 2FA). 704 tools are exposed this way.
• Developer API: /api-gateway/v1/*, authenticated with an X-API-Key header. See /developers.
• Slack / Microsoft Teams: see the ChatOps page in the app.

Running it locally

Requirements: Python 3.11+, PostgreSQL (or let the tests use SQLite), and an Anthropic and/or Moonshot (Kimi) API key.

python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

export SECRET_KEY=change-me
export ANTHROPIC_API_KEY=sk-ant-...      # Claude tiers
export MOONSHOT_API_KEY=...              # Kimi tiers
export DATABASE_URL=postgresql://user:pass@localhost:5432/devops_agent   # required in production

python app.py                            # http://127.0.0.1:5000

New tables and columns are created automatically at startup. For a local Postgres and Redis, run docker compose up db redis after setting POSTGRES_PASSWORD in .env (compose refuses to start without it). The compose file's app service has no Dockerfile to build from, so run the app itself with python app.py.

Cloud and integration credentials are not read from environment variables. Each user adds their own in Settings once the app is running.

Command-line mode

python main.py interactive     # chat in the terminal
python main.py ask "list pods in the default namespace"
python main.py tools           # list tools

The CLI has no signed-in user, so AWS, Azure and GCP tools refuse to run there (they only use per-user credentials). Kubernetes works with your local kubeconfig when you set K8S_USE_LOCAL_KUBECONFIG=true (on the hosted site it reaches the user's EKS clusters instead); local Git, Docker and Terraform work, and integration tools work when you pass a token explicitly. See docs/USER_GUIDE.md.

Tests

pytest -q

Tests run against in-memory SQLite, so no database is needed. Two network-safety tests need public DNS and fail in fully offline sandboxes.

Project layout

app.py              Flask app: pages, chat endpoints, startup
models.py           SQLAlchemy models
src/agent/          Agent loop, safety classification, tool selection, guardrails
src/tools/          Tool modules (aws_tools.py, kubernetes_tools.py, ...)
services/           Business logic (credentials, workflows, approvals, billing, ...)
routes/             Blueprints for feature pages and APIs
api/v1/             Deprecated REST API (/api/v1); its handlers also serve /api-gateway/v1
templates/          Jinja2 pages
tests/              pytest suite
docs/               Guides and references

Contributors: read CLAUDE.md first. It records the security rules and invariants the code depends on (per-user isolation, SSRF guards, risk classification, guardrails), and CLAUDE_RUNBOOK.md records past incidents and fixes.

Removed features

These were removed in October 2026 and are no longer part of the app: Optimize, Monitoring (alert rules), Incidents (and playbooks), Analysis (root-cause analysis) and Predictive Scaling. Incident work now happens in chat, with the PagerDuty and OpsGenie tools and the Incident Response workflows.

Support

• In the app: the chat, or /docs
• Email: support@devopsagent.io
• Security issues: security@devopsagent.io (see /security)