MCP
The MCP Server drives the rest of 10x from a chat session. Install it into Claude Desktop, Claude Code, or Cursor; ask cost questions, generate per-app Helm values, and apply filter caps without leaving the conversation.
Overview
Do I need MCP to use 10x
No. Every 10x app (Reporter, Receiver, Retriever) has standalone install/deploy docs you can follow manually. MCP just makes adoption faster by knowing your environment and generating tailored configs.
If you prefer a manual workflow, skip MCP entirely and follow each app's Deploy page.
What it does
How does MCP know about my cluster
When you ask MCP to discover your environment, it reads your kubeconfig (read-only, same credentials as kubectl get) to identify:
- Forwarder DaemonSets (Fluent Bit, Fluentd, Datadog Agent, OTel Collector, etc.)
- Node counts and pod topology
- Existing logging destinations (Splunk HEC endpoints, Elasticsearch clusters, CloudWatch log groups)
It then proposes a deployment plan tailored to your stack. See Tools for the per-tool reference.
Does MCP apply changes to my infrastructure
No. MCP is an advisor, not an actor:
- k8s discovery is read-only (
kubectl getequivalents only) - Generated configs are files, not applied changes. MCP writes
my-reporter.yaml; you review andhelm install. - Filter/compact suggestions are diff-reviewable. MCP proposes mute file entries; you commit.
- Dependency checks call the log analyzer's API in-process when creds are in the env (read-only, never POST/PUT/DELETE), and fall back to copy-paste bash when they aren't.
You stay the operator. MCP never writes to your cluster, your git repo, or your log analyzer.
Security
What data does the MCP Server see
- API calls to your metrics backend (pre-aggregated metrics, no log content). The default hosted endpoint (prometheus.log10x.com) is optional; you can point at your own Prometheus or run air-gapped.
- kubeconfig read access (only what
kubectl getwould surface) - Your log analyzer credentials (for dependency checks; read-only, local-only)
No log content ever leaves your machine. No data is cached server-side.
Where does the MCP Server run
As a local subprocess of your AI assistant (Claude Desktop, Code, Cursor). It starts when the assistant connects to the MCP and stops when the assistant closes. No cloud hosting, no Lambda, no API Gateway.