devops MCP server
Commit Nov 13, 2026Allows Claude to read data from Kubernetes, Jenkins, GitHub Actions, AWS, Azure, GCP, ServiceNow, Slack, and Teams to assist on-call engineers.
Also known as gerardrecinto/devops-mcp
- Stars
- 1
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 13, 2026
- License
- MIT
What is devops MCP?
devops allows Claude to read data from Kubernetes, Jenkins, GitHub Actions, AWS, Azure, GCP, ServiceNow, Slack, and Teams to assist on-call engineers.
Summary generated from the project's README
About devops MCP server
On-call engineers can investigate system failures and correlate events across multiple platforms without switching between different dashboards. This server provides a unified way for Claude to read data from Kubernetes, Jenkins, GitHub Actions, AWS, Azure, GCP, ServiceNow, Slack, and Teams. It is designed to help identify root causes by merging information from CI/CD pipelines, cloud infrastructure, and incident management tools into a single view.
Users can run this locally via Claude Desktop or Claude Code using Python 3.11 or Docker. Configuration is managed through a .env file located at ~/.devops-mcp/.env, where users provide credentials for their specific services. The server is read-only by default, meaning it cannot delete or restart resources. The only exception is sending messages to Slack or Teams, which requires the DEVOPS_MCP_ENABLE_WRITES environment variable to be explicitly enabled.
What sets this tool apart is the incident_timeline capability, which joins changes from various sources to find suspects that landed just before a failure. It also includes safety features like an audit log and the ability to use Joltrin to prevent the AI from posting messages to Slack or Teams until it has verified real data from the system.
What you can do with it
- 1Correlate recent system changesAsk the AI to find what changed and what broke in the last few hours across different platforms.
- 2Debug failing Kubernetes podsIdentify which pods are crashing and retrieve their specific logs or error events.
- 3Review CI/CD pipeline statusDetermine if specific Jenkins builds or GitHub Action workflows have failed.
Written from the project's README. Every use case is checked against a quote from it.
Configuration
DEVOPS_MCP_DEMODEVOPS_MCP_ORDER
| Name | Description | |
|---|---|---|
| DEVOPS_MCP_DEMO | — | |
| DEVOPS_MCP_ORDER | — |
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablegerardrecinto/devops-mcp is public and reachablegithub.com/gerardrecinto/devops-mcp ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/gerardrecinto/devops-mcp ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubdevops-mcp
devops-mcp lets Claude read from your Kubernetes, Jenkins, GitHub Actions, AWS, Azure, GCP, ServiceNow, Slack and Teams, so on-call engineers can ask one question instead of checking six dashboards.
New here? Claude can look at your servers and build systems for you and explain what it finds, but it cannot change them.
It is an MCP server. MCP (Model Context Protocol) is the way Claude Code and Claude Desktop call outside tools. Site with the tool browser: https://gerardrecinto.github.io/devops-mcp/
Try it in 60 seconds
You need Python 3.11 or later, pipx on your PATH, and Claude Code installed to register the server.
pipx install "git+https://github.com/gerardrecinto/devops-mcp.git"
claude mcp add devops-mcp -- devops-mcp
devops-mcp --check
--check prints which services have local configuration and what to do next. It does not make network requests or verify credentials. Then ask Claude:
Which pods are failing in my cluster?
With no other setup it uses your current kubeconfig (~/.kube/config), so the answer is about your real cluster. A service you have not set up returns a plain "not configured" error.
This installs from GitHub on purpose. The devops-mcp package on PyPI is a different project, and this one is not published there yet.
Example
Illustrative only. The pod names, incident number and times are made up to show the flow. They are not output from a real system.
You: "The API deployment has been stuck for 20 minutes. What's going on?"
Claude:
[k8s_get_deployment_status("api-server", "production")]
desired: 3 ready: 1 updated: 3 strategy: RollingUpdate
[k8s_get_failing_pods("production")]
api-server-7d9f8b-xk2p9 CrashLoopBackOff restarts: 14
[k8s_pod_logs("api-server-7d9f8b-xk2p9", tail_lines=30)]
[09:10:03] FATAL postgresql://db-prod:5432: Connection refused
[k8s_get_events("production", event_type="Warning")]
OOMKilling Pod/db-prod-0 "Memory limit exceeded" count: 3
Claude: db-prod-0 was OOMKilled, so the API pods cannot reach the database.
Connect more systems
Copy .env.example to ~/.devops-mcp/.env and fill in only the services you use. The server reads that file on start, then restart your client. Run devops-mcp --check again to check that the configuration is present. An actual tool call verifies access.
Azure, GCP and Slack need extra packages:
pipx install --force "mcp-devops[azure,gcp,slack] @ git+https://github.com/gerardrecinto/devops-mcp.git"
To see the output shape before connecting anything, add the demo flag. Every response is then labelled as sample data and is not real information about your systems. Remove the flag once you connect real systems.
claude mcp add devops-mcp -e DEVOPS_MCP_DEMO=1 -- devops-mcp
Tracing and SLO helpers (optional)
Two small add-ons that the server does not turn on by itself.
devops_mcp.tracing: OpenTelemetry spans around a tool call. Install withpip install "mcp-devops[otel]", callconfigure_tracing(), then wrap a function with@traced("tool_name"). Spans record the tool name and argument names, never argument values. A sample collector config is indocs/observability/otel-collector.yaml.devops_mcp.slo: multi-window burn-rate alerts (the 14.4x, 6x, 3x and 1x pattern from the Google SRE Workbook). Try it:python -m devops_mcp.slo --slo 0.999 --window 1h=200/10000 --window 5m=20/1000.
Tools
33 tools. The full list with arguments is in the tool browser, and the tables are below.
| Service | Tools |
|---|---|
| Incident timeline | 1 |
| Kubernetes | 5 |
| Jenkins | 3 |
| GitHub Actions | 2 |
| AWS | 4 |
| Azure | 4 |
| GCP | 4 |
| ServiceNow | 3 |
| Slack and Teams | 6 |
| Server health | 1 |
Optional: no announcement before a real look (with Joltrin)
slack_post_message and teams_post_message can tell a whole channel something false if an assistant posts before it has looked at anything. With Joltrin the server can refuse a post until this session has read real data from a system.
It needs a Joltrin build with the gate command (in Joltrin's master, not yet in a release: go install github.com/sharedcode/joltrin/v5/cmd/sop-mcp-server@master). Copy examples/order.json and examples/runbooks.json next to each other, then start the server with the file named in the environment:
DEVOPS_MCP_ORDER=$PWD/examples/order.json DEVOPS_MCP_ENABLE_WRITES=1 devops-mcp
A post then returns order_not_satisfied until one of the tools listed under record (for example k8s_get_failing_pods) has returned real data in the same server session.
What to know:
- Off unless
DEVOPS_MCP_ORDERis set. Nothing else changes. - Only real data counts. An error, or sample data from demo mode, does not.
- If the file or
sop-mcp-servercannot be used, the two post tools are refused and say why. Read tools keep working. - A session is one server process, or the name in
DEVOPS_MCP_ORDER_SESSION. A long-running server keeps "already looked" until it restarts. - It checks that a read happened, not that its result was good news.
What it does and does not do
- It is read-only by default. Read-only means no tool applies, deletes, restarts or triggers anything.
- The only writes are Slack and Teams messages, and they are refused unless you set
DEVOPS_MCP_ENABLE_WRITES=1. - Tool calls are logged to
~/.devops-mcp/audit.logwhen audit storage is available, with secret-like values redacted. Audit failures produce a warning and leave the server usable. - It reports what the APIs return. Claude's explanation of a root cause is a hypothesis to check, not a verdict.
- It does not create or close ServiceNow records, run
kubectl apply, or start builds.
More detail
Incident timeline
| Tool | What it answers |
|---|---|
incident_timeline |
What changed, and what broke, in the last N hours? One ordered view across ServiceNow, Kubernetes, GitHub Actions and Jenkins, with the changes that landed just before the first failure listed as suspects |
Suspects are matched by time proximity (a change within 2 hours before the first failure) and the response says so. A source that is unconfigured, in demo mode or failing is reported as skipped and never filled in with sample data.
Kubernetes
| Tool | What it answers |
|---|---|
k8s_get_failing_pods |
What pods are crashing right now? |
k8s_pod_logs |
What is this pod printing? |
k8s_describe_pod |
Full state: conditions, restart counts, events |
k8s_get_events |
What Warnings has the cluster emitted? |
k8s_get_deployment_status |
Is this rollout progressing or stuck? |
Jenkins
| Tool | What it answers |
|---|---|
jenkins_get_build_status |
Did build #N succeed, and how long did it take? |
jenkins_get_build_log |
Last N lines of a build |
jenkins_list_failing_jobs |
Which jobs are currently red? |
GitHub Actions
| Tool | What it answers |
|---|---|
gh_get_workflow_run |
Status and conclusion for a specific run |
gh_list_failed_runs |
Which workflows have failed in the last 24 hours? |
AWS
| Tool | What it answers |
|---|---|
aws_get_cloudwatch_logs |
What did this log group emit? |
aws_get_metric |
Current CloudWatch metric value |
aws_describe_s3_bucket |
Versioning, lifecycle, encryption config |
aws_get_ecs_service_status |
Running vs desired vs pending |
Azure
| Tool | What it answers |
|---|---|
azure_get_monitor_logs |
KQL query against Log Analytics |
azure_get_metric |
Azure Monitor metric |
azure_get_aks_node_status |
AKS node pool state |
azure_get_resource_health |
Is this Azure resource healthy? |
GCP
| Tool | What it answers |
|---|---|
gcp_get_logs |
Cloud Logging filter query |
gcp_get_metric |
Cloud Monitoring metric |
gcp_get_gke_cluster_status |
GKE cluster and node pool status |
gcp_get_cloud_run_status |
Cloud Run revision and traffic split |
ServiceNow
| Tool | What it answers |
|---|---|
snow_list_incidents |
Open incidents right now |
snow_get_incident |
Full detail on INC0012345 |
snow_get_change_requests |
Scheduled or in-flight changes |
Slack / Teams
| Tool | What it does |
|---|---|
slack_post_message |
Post to a channel (needs DEVOPS_MCP_ENABLE_WRITES=1) |
slack_get_channel_history |
What was said in #incidents recently? |
slack_search_messages |
Search Slack messages across channels |
teams_post_message |
Post an alert card via webhook (needs DEVOPS_MCP_ENABLE_WRITES=1) |
teams_get_channel_messages |
Recent messages in a Teams channel (Graph API) |
teams_list_channels |
Channels in a Teams team (Graph API) |
Server
| Tool | What it answers |
|---|---|
devops_mcp_health |
Which integrations are configured, and are demo mode and writes on? |
Tool failures come back as {"error", "message", "hint"} instead of a stack trace.
The image includes every integration. It speaks MCP over stdio, so the client starts it:
docker pull ghcr.io/gerardrecinto/devops-mcp:latest
claude mcp add devops-mcp -- docker run -i --rm --env-file ~/.devops-mcp/.env \
-v ~/.kube:/home/devops-mcp/.kube:ro ghcr.io/gerardrecinto/devops-mcp:latest
Add to ~/Library/Application Support/Claude/claude_desktop_config.json, then restart:
{
"mcpServers": {
"devops-mcp": {
"command": "devops-mcp"
}
}
}
Credentials go in ~/.devops-mcp/.env (see .env.example). To try sample data first, add "env": { "DEVOPS_MCP_DEMO": "1" } and remove it later.
A first draft of "wrap the Kubernetes API in an MCP tool" is easy. The parts that make it safe to run against production are not, and each one is tested here:
- One timeline instead of six dashboards.
incident_timelinemerges changes, CI failures, cluster warnings and incidents into one ordered view and names the changes that landed shortly before things broke. That cross-system join is the part a single-service wrapper does not give you. - It never invents data. An unconfigured service returns a plain "not configured" error. Sample data is opt-in (
DEVOPS_MCP_DEMO=1) and every response is labelled, so an assistant cannot report a made-up OOMKill or a Slack message that was never sent. - It finds the failures that matter. A crash-looping pod reports phase
Running; the reason sits on the container state. The pod check reads container waiting and terminated reasons, so CrashLoopBackOff and ImagePullBackOff are caught, not justPending. - Read-only by default. No tool applies, deletes, restarts or triggers anything. The only writes are Slack and Teams messages, and those are refused unless the operator sets
DEVOPS_MCP_ENABLE_WRITES=1. - Prompt-injection aware. A webhook URL supplied by the model is ignored unless its host is in
DEVOPS_MCP_TEAMS_WEBHOOK_ALLOWLISTand is an https Microsoft endpoint, so injected text in a log line cannot make the server POST channel content to an arbitrary host. - A private audit trail. When audit storage is available, each tool call is logged as one JSON line in
~/.devops-mcp/audit.log(mode 0600, rotated at 10 MB). Secret-like parameters are redacted, message tools log the destination and a SHA-256 of the body instead of the body, and an audit initialization or write failure is logged without stopping the server. - Bounded output. Pod logs are capped by lines and bytes (a cut is marked
[truncated ...]), and Jenkins console logs are streamed so a multi-hundred-MB build log cannot exhaust memory. - Partial installs work. Each integration loads on its own; one broken module never stops the rest.
See docs/architecture.md for the full protocol flow, tool registration, and audit logging diagrams.
+-----------------------------------------------------------------------+
| Claude (LLM) |
| "Why is production down?" / "Which jobs are failing right now?" |
+-----------------------------------+-----------------------------------+
| MCP Protocol (stdio)
v
+-----------------------------------------------------------------------+
| devops-mcp server |
| |
| Kubernetes | Jenkins/GHA | AWS | Azure |
| pods/events | build logs | CW logs | Monitor/AKS |
| deployment | failing jobs | metrics/ECS | resource health |
| | | | |
| GCP | ServiceNow | Slack | Teams |
| Cloud Logs | incidents | post/search | post/messages |
| GKE/Run | changes | history | channels |
| |
| Audit log (~/.devops-mcp/audit.log) |
+-----------------------------------------------------------------------+
Each tool module (Kubernetes, Jenkins, AWS, etc.) loads independently. If a module fails to import, the server logs the error and continues. Every other module still registers its tools normally. The version is logged at startup.
A service without credentials answers with a "not configured" error instead of guessing. Set DEVOPS_MCP_DEMO=1 to get labelled sample data instead.
Infrastructure tools are read-only. Claude can observe and explain, not modify. The Slack and Teams post tools are the only writes, and they are off until you set DEVOPS_MCP_ENABLE_WRITES=1.
| Action | Available |
|---|---|
| Read pod logs, events, deployment status | yes |
| Read build logs and pipeline status | yes |
| Read CloudWatch / Azure Monitor / Cloud Logging | yes |
| Read ServiceNow incidents and changes | yes |
| Post to Slack / Teams | only with DEVOPS_MCP_ENABLE_WRITES=1 |
kubectl apply / trigger builds / modify resources |
no |
| Create or close ServiceNow records | no |
To let the model pick a Teams webhook per call, list the allowed hosts in DEVOPS_MCP_TEAMS_WEBHOOK_ALLOWLIST (comma-separated). Otherwise only TEAMS_WEBHOOK_URL is used.
Credentials are environment variables only. Each service is independently optional.
When audit storage is available, tool calls are appended to ~/.devops-mcp/audit.log with secret-like parameters redacted.
For the commercial path, launch angles, and first-customer profile, see docs/go-to-market.md.
pip install -e ".[dev]"
pytest tests/ -v --cov=devops_mcp
- Grafana alert history
- PagerDuty incident lookup
- Terraform state drift detection
- ArgoCD sync status
- SSE transport
For the commercial path and first-customer profile, see docs/go-to-market.md.
License
MIT
Optional paid help: scoped reviews and setup.
devops MCP questions
Is this server read-only?
Yes, it is read-only by default. No tool can apply, delete, restart, or trigger changes, except for Slack and Teams messages which require explicit enablement.
How can I see sample data without connecting my real services?
You can use the demo flag to see the output shape using sample data labeled as such.
Does it support Kubernetes configuration?
Yes, it uses your existing local kubeconfig file to access your real cluster.
Where are tool calls logged?
Tool calls are recorded in a private audit log located in your home directory.