🗣️ KubeWhisper

An AI-DevOps MCP server — talk to your infrastructure in plain English, safely.

KubeWhisper is a Model Context Protocol (MCP) server that gives an LLM like Claude read-only-by-default access to a live environment, so you can ask:

"Which pods are crash-looping in staging?" "What's the p95 latency right now?" "Did the last GitHub Actions run on main pass?" "Scale the web deployment to 5 — but show me the preview first."

…and the model answers grounded in your actual cluster, metrics, and CI — not generic guesses.

MCP crossed ~97M monthly SDK downloads in March 2026 and is now adopted by Anthropic, OpenAI, Google, Microsoft and Amazon. Plenty of people use MCP servers; far fewer have built one for real DevOps work. This is a small, honest, safety-first example of how.

✨ Tools

Related MCP server: InfraClaude

🌐 Multi-cluster + error-context routing

Drop a clusters.json next to the server (or point KUBEWHISPER_CLUSTERS at it). Each cluster maps to a kubeconfig context, a Prometheus URL, and routing hints:

{
  "default_cluster": "prod-mumbai",
  "clusters": [
    { "name": "prod-mumbai", "kube_context": "gke_proj_asia-south1_prod",
      "prometheus_url": "http://prometheus.prod-mumbai:9090",
      "match": { "keywords": ["prod","mumbai","live","p1"], "namespaces": ["streaming","ingest"] } },
    { "name": "staging", "kube_context": "gke_proj_asia-south1_staging",
      "prometheus_url": "http://prometheus.staging:9090",
      "match": { "keywords": ["staging","qa"], "namespaces": ["streaming-staging"] } }
  ]
}

Now the agent can route by the error context: given an alert like "P1: transcode pods OOMing on the Mumbai live stream", it calls resolve_cluster, matches mumbai / live / transcode → prod-mumbai, and targets that cluster's kube context + Prometheus automatically. Every cluster-scoped tool also accepts an explicit cluster argument. With no clusters.json, it falls back to your current kube context. See clusters.example.json.

💬 Slack bot

Two ways to use Slack:

1. Outbound notes from any MCP client — the send_slack_message tool (set SLACK_BOT_TOKEN or SLACK_WEBHOOK_URL).

2. A full Slack bot (slack_bridge.py) — @mention it in a channel ("@KubeWhisper which pods are failing in prod?") and it runs a Claude tool-use loop over these tools, routes to the right cluster, and replies in-thread.

pip install -r requirements-slack.txt
export ANTHROPIC_API_KEY=...      # console.anthropic.com
export SLACK_BOT_TOKEN=xoxb-...   # scopes: app_mentions:read, chat:write
export SLACK_APP_TOKEN=xapp-...   # Socket Mode
python slack_bridge.py

The bot is read-only by default — scale_deployment stays preview-only unless an operator sets KUBEWHISPER_ALLOW_WRITES=true on the bot process.

🔐 Safety model (the whole point)

1. Read-only by default. The single write tool is off unless KUBEWHISPER_ALLOW_WRITES=true.

2. Preview, then confirm. A write with confirm=False only describes the change.

3. Audit trail. Every write intent — refused, previewed, or applied — is appended to a JSONL log.

4. No destructive verbs. delete / drain / cordon are deliberately not implemented in v0.

See docs/architecture.md for the diagram.

🚀 Quickstart

# 1. Install
python -m venv .venv && source .venv/bin/activate   # (Windows: .venv\Scripts\activate)
pip install -r requirements.txt

# 2. Point at a cluster + Prometheus (a local `kind` cluster is perfect)
export PROMETHEUS_URL=http://localhost:9090
#   KUBEWHISPER_ALLOW_WRITES stays unset → read-only

# 3. Run it
python -m kubewhisper.server

Use it from Claude Desktop

Copy the block in config.example.json into your claude_desktop_config.json, fix the cwd path, restart Claude Desktop, and ask it about your cluster.

🧪 Tests

pip install pytest
pytest -q          # guardrail tests prove writes are refused unless explicitly enabled

CI runs lint + import check + guardrail tests on every push (see .github/workflows/ci.yml).

🗺️ Roadmap

• v0.1 — read-only tools + guarded scale + audit (this release)

• v0.2 — Terraform plan (read-only) tool + deploy "what-changed" diff

• v0.3 — eval suite proving destructive-op refusal + OPA policy layer over writes

• v0.4 — one-command docker run, publish to an MCP registry

⚠️ Disclaimer

A learning / portfolio project. Run it against clusters you own. Keep it in read-only mode unless you fully understand the write path. Built independently, on personal infrastructure.

📄 License

MIT — see LICENSE.

Built by Abdulhussain Kanchwala · Portfolio · LinkedIn · GitHub