Domain Agents›Kubernetes Agent›MCP Tool Servers

MCP servers

TalkOps Kubernetes Agent (k8s-autopilot) uses the Model Context Protocol (MCP) to connect specialist operators to Kubernetes clusters, GitOps controllers, and observability backends. The agent ships with 11 built-in MCP servers (220 tools total) and lets you register custom third-party or internal MCP servers via the Settings UI or .mcp.json files.

Built-in MCP servers

MCP server	Assigned operator	Tools	Primary capabilities
kubernetes-mcp-server	K8s Operator	32	Workload CRUD, pod logs & exec, events, node diagnostics, RBAC, multi-cluster context switching
helm-mcp-server	Helm Operator	22	Chart discovery, dry-run validation, install, upgrade, rollback, uninstall, JSON schema checks
argocd-mcp-server	App Operator	29	Application onboarding, sync, rollback, health/diff inspection, project and repo management
argo-rollout-mcp-server	App Operator	18	Canary and blue-green promotions, abort/retry, AnalysisTemplate verification
traefik-mcp-server	App Operator	11	IngressRoute generation, weighted traffic splitting, middlewares, TLS, NGINX migration
prometheus-mcp-server	Observability	28	PromQL instant/range queries, ServiceMonitor & PrometheusRule CRUD, cardinality analysis
alertmanager-mcp-server	Observability	14	Active alert triage, silence creation with blast-radius preview, routing tree audits
opentelemetry-mcp-server	Observability	19	Collector CRD provisioning, workload auto-instrumentation, OTLP pipeline validation
loki-mcp-server	Observability	9	LogQL stream and metric queries, label discovery, trace-to-log correlation
tempo-mcp-server	Observability	16	TraceQL search, critical-path waterfall analysis, RED metrics, service topology graphs
github-mcp-server	Helm / Core	12	Repository branch management, Helm chart commits, pull request creation

Register a custom MCP server

Via the Settings UI

Navigate to Settings → MCP Servers and click + Add Server. Specify the server name, transport (stdio, http, or sse), command or URL, and environment variables. Newly added servers become available immediately without restarting containers.

Via .mcp.json

Define servers in .mcp.json using the standard MCP configuration schema:

.mcp.json

{
  "mcpServers": {
    "postgres-inspector": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres"],
      "env": {
        "DATABASE_URL": "${POSTGRES_READONLY_URI}"
      }
    },
    "internal-cmdb": {
      "url": "http://cmdb-mcp.internal:8080",
      "transport": "http",
      "headers": {
        "Authorization": "Bearer ${CMDB_API_TOKEN}"
      },
      "allowed_tools": ["get_*", "list_*"],
      "disabled_tools": ["delete_*"]
    }
  }
}

Field	Type	Description
command	string	Executable for stdio transport (npx, uvx, python, docker)
args	list[string]	Command-line arguments passed to command
env	object	Environment variables passed to the subprocess (supports ${VAR} and ${VAR:-default})
url	string	Endpoint URL for remote http or sse servers
transport	string	stdio (default when command is set), http, or sse
headers	object	HTTP headers sent on remote http or sse connections
allowed_tools	list[string]	Glob allowlist of tool names to expose
disabled_tools	list[string]	Glob denylist of tool names to hide (evaluated before allowed_tools)
enabled	bool	Set false to disable a server without removing its configuration

Configuration discovery order

k8s-autopilot merges MCP server definitions from five sources:

1. Settings UI database (db) — Highest precedence.
2. Global user files (~/.agents/.mcp.json or ~/.k8s_autopilot/.mcp.json).
3. Project root files (.mcp.json or .k8s_autopilot/.mcp.json).
4. Vertical plugin configs (plugins/<name>/.mcp.json).
5. Operator-scoped configs (built_in_subagents/<operator>/.mcp.json) — Isolated to that specific operator and connected Just-In-Time (JIT).

Tool naming and security profiling

Every discovered MCP tool is registered under the namespace mcp__<server>__<tool> (for example, mcp__kubernetes__get_pods) and classified into one of four security tiers:

Tier	Classification	Governance behavior
Tier 1 — Read-only	readOnlyHint=true or inspection verbs (get_, list_, describe_, query_)	Executes automatically
Tier 2 — Low-impact	Dry-runs or metadata annotations (dry_run_, annotate_)	Auto mode evaluates; Manual mode asks
Tier 3 — Mutating	Live state changes (install_, scale_, apply_, sync_)	Requires HITL approval in Manual and Auto modes
Tier 4 — Destructive	destructiveHint=true or deletion verbs (delete_, uninstall_, drain_)	Requires HITL approval with Critical risk warning

Project-level .mcp.json files committed to Git are fingerprinted with SHA-256 on first approval; if a pull request modifies .mcp.json, k8s-autopilot prompts for re-approval before spawning the server.