Sponsored by Deepsite.site

Talos Linux Mcp Server

Created By
Nosmohta month ago
MCP server for Talos Linux cluster management via the native gRPC API
Content

talos-mcp

CI Release Go Reference codecov Go Report Card OpenSSF Scorecard License

An MCP server that exposes Talos Linux cluster management to AI agents (Claude Code, OpenAI Codex, and any MCP-compatible client). Instead of pasting talosctl output into chat, the agent calls structured tools that return machine-readable JSON directly from the Talos gRPC API — zero token cost for intermediate output.

Connects to your cluster via the native Talos gRPC API using the same mTLS credentials as talosctl (~/.talos/config).

Installation

Via npm (no Go required, Linux/macOS, amd64/arm64):

npx talos-mcp

Download binary (Linux/macOS, amd64/arm64):

Download the latest release from GitHub Releases, extract, and place the binary in your $PATH.

Build from source (requires Go 1.21+):

git clone https://github.com/Nosmoht/talos-mcp-server
cd talos-mcp
go build -o talos-mcp ./cmd/talos-mcp

Configuration

Reads ~/.talos/config by default (the same file talosctl uses). Override via environment variables:

VariableDefaultDescription
TALOSCONFIG~/.talos/configPath to talosconfig file
TALOS_CONTEXTactive contextContext name to use
TALOS_ENDPOINTSfrom configComma-separated endpoint overrides
TALOS_MCP_READ_ONLYfalseSet to true to disable all mutating tools at startup
TALOS_MCP_HTTP_ADDR(unset)If set (e.g. :8080), serve Streamable HTTP instead of stdio
TALOS_MCP_AUTH_TOKEN(unset)Required bearer token when HTTP mode is active
TALOS_MCP_ALLOWED_NODES(unset)Comma-separated IPs, hostnames, and CIDR ranges permitted as tool targets. Unset allows all.
TALOS_MCP_ALLOWED_PATHS(all)Comma-separated path prefixes allowed for talos_read_file and talos_list_files (e.g. /etc,/proc)
TALOS_MCP_SKIP_VERSION_CHECKfalseSet to true to bypass upgrade path validation (e.g. for factory images or custom tags)
TALOS_MCP_RATE_LIMIT10HTTP mode: token-bucket refill rate (requests/second, float)
TALOS_MCP_RATE_BURST20HTTP mode: token-bucket burst capacity (int)
TALOS_MCP_MAX_BODY_SIZE4194304HTTP mode: max POST request body size in bytes (4 MiB default)
TALOS_MCP_MAX_CONCURRENT20HTTP mode: max concurrent POST handlers (fail-fast 503 on overload)
TALOS_MCP_SUBSCRIPTION_RATE1sMinimum interval between delivered resources/updated notifications per (session, URI) pair (Go duration, e.g. 500ms)
TALOS_MCP_SUBSCRIPTION_BURST3Initial notification burst per (session, URI) before the rate kicks in

Compatibility

This server is tested against Talos Linux v1.9.x through v1.12.x.

talos-mcpTalos Linuxmachinery SDK
v0.x (current)v1.9.0 – v1.12.xv1.12.6

The server logs a startup warning if the connected cluster's Talos version is outside the tested range. All 19 gRPC methods used have been stable since Talos v1.9.

Upgrade path validation

The talos_upgrade tool validates that the target version follows Talos's supported upgrade path — at most one minor version at a time (e.g. v1.11.x → v1.12.x). Upgrades that skip minor versions are rejected with an error.

If your image uses a custom or factory tag (e.g. factory.talos.dev/... or :latest) the tag cannot be parsed and validation is skipped automatically. To bypass validation explicitly, set TALOS_MCP_SKIP_VERSION_CHECK=true.

Client Setup

Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "talos": {
      "command": "npx",
      "args": ["-y", "talos-mcp"]
    }
  }
}

Or globally in ~/.claude.json under "mcpServers". If you prefer a local binary, replace "command": "npx" with the path to the binary.

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "talos": {
      "command": "npx",
      "args": ["-y", "talos-mcp"]
    }
  }
}

OpenAI Codex

Add to .codex/config.toml (project) or ~/.codex/config.toml (global):

[mcp_servers.talos]
command = "npx"
args = ["-y", "talos-mcp"]

[mcp_servers.talos.env]
TALOSCONFIG = "/path/to/talosconfig"

Generic MCP client

The server speaks the MCP protocol over stdio:

./talos-mcp

Tools

Read-only

ToolDescription
talos_resource_definitionsList all available resource types and their aliases. Call this first to discover what can be queried.
talos_getGet or list any COSI resource by type (e.g. MachineStatus, Member, NodeAddress, Service).
talos_versionGet Talos version info from target nodes.
talos_servicesList all Talos services and their current state (running, stopped, health).
talos_containersList containers in a namespace (default: k8s.io for Kubernetes containers).
talos_processesList running processes on target nodes.
talos_healthCheck cluster health (etcd, Kubernetes API, node readiness). Supports control_plane_nodes / worker_nodes override.
talos_logsFetch recent service logs (last N lines, no follow).
talos_dmesgRead kernel ring buffer messages.
talos_eventsFetch recent Talos runtime events (service changes, config changes).
talos_etcdQuery etcd cluster: members (default) or status.
talos_list_filesList files and directories on a node filesystem.
talos_read_fileRead file contents from a node filesystem.

Mutating

These tools modify cluster state and have explicit safety guards.

ToolDescriptionGuards
talos_service_actionStart, stop, or restart a Talos service (note: restarting etcd is not supported by the Talos API).confirm=true required
talos_rebootReboot target nodes. Supports mode: default, powercycle, force.confirm=true required; nodes must be explicit
talos_upgradeUpgrade Talos on target nodes. Supports preserve (default true), stage, force, reboot_mode.confirm=true required; nodes and image required
talos_rollbackRoll back the last upgrade on target nodes.confirm=true required; nodes must be explicit
talos_patch_configApply a machine config patch (JSON or YAML strategic merge).dry_run defaults to true; confirm=true required when dry_run=false

All tools accept an optional nodes field (list of node IPs or hostnames). When omitted, the active context from talosconfig is used.

Prompts

PromptDescription
diagnose-nodeGuided diagnosis workflow for a single node.
investigate-etcdFocused investigation of an etcd cluster anomaly.
debug-serviceService-specific diagnostic workflow (kubelet, containerd, etcd, …).
pre-upgrade-checklistPre-flight verification before a Talos upgrade.
apply-configGuided flow for applying a machine config patch (registered only when TALOS_MCP_READ_ONLY is unset).

Resources and Subscriptions

The server exposes Talos COSI resources as MCP resources:

  • talos://cluster/version — static cluster version info.
  • talos://cluster/resource-definitions — discover resource types.
  • talos://{node}/resource/{namespace}/{type}[/{id}] — list or get COSI resources on a specific node.

MCP clients that implement resources/subscribe (Claude Desktop, Cursor) receive notifications/resources/updated whenever the underlying resource changes — no polling required. Subscriptions are backed by the Talos COSI Watch / WatchKindAggregated streams and honour the same TALOS_MCP_ALLOWED_NODES allowlist as reads.

Subscribable resource types (canonical names):

  • MachineStatuses.runtime.talos.dev (MachineStatus)
  • Members.cluster.talos.dev (Member)
  • NodeAddresses.net.talos.dev (NodeAddress)
  • Services.v1alpha1.talos.dev (Service)

Aliases resolve to the canonical type before the allowlist check, so a client subscribing to talos://{node}/resource/runtime/ms/... (alias for MachineStatus) succeeds. Other COSI types reject with resource type %q is not subscribable. Static talos://cluster/* URIs are not subscribable (no COSI backing).

Delivery is rate-limited per (session, URI) via TALOS_MCP_SUBSCRIPTION_RATE / TALOS_MCP_SUBSCRIPTION_BURST; over-rate events are dropped and the client re-reads the resource to catch up. The initial Bootstrapped event is intentionally not forwarded — the client is expected to call resources/read once after subscribe for initial state.

Security Model

Trust Boundaries

MCP Client (Claude Code / Codex)
        │  stdio / JSON-RPC
   talos-mcp  ◄── reads TALOSCONFIG (~/.talos/config)
        │  gRPC + mTLS
  Talos API (each node)
    Node OS

Data flow warning: Tool responses flow directly into the LLM's context window and are sent to the LLM provider. Anything a tool returns — node IPs, hostnames, service configurations, kernel logs, file contents — becomes part of the prompt sent over the network. Do not use this server with clusters containing data you would not be comfortable sending to your LLM provider.

Talos RBAC is server-side enforced. The credentials in your talosconfig determine what operations are permitted on each node. talos-mcp cannot bypass Talos RBAC — a request that the API rejects will fail with an error, not silently succeed.

Tool Classification and Minimum Required RBAC Role

ToolRBAC minimum
talos_resource_definitions, talos_get, talos_version, talos_services, talos_containers, talos_processes, talos_health, talos_logs, talos_dmesg, talos_events, talos_list_files, talos_read_fileos:reader
talos_etcd, talos_service_action, talos_reboot, talos_upgrade, talos_rollbackos:operator
talos_patch_configos:admin

Safety Mechanisms

MechanismHow it works
Read-only modeTALOS_MCP_READ_ONLY=true registers only read-only tools at startup; mutating tools are never exposed to the LLM
Path allowlistTALOS_MCP_ALLOWED_PATHS=/etc,/proc restricts talos_read_file and talos_list_files to specified prefixes
Confirm gatesAlways require confirm=true: talos_service_action, talos_reboot, talos_upgrade, talos_rollback, talos_reset. Require confirm=true when dry_run=false: talos_patch_config, talos_apply_config. All enforced server-side.
Preserve defaulttalos_upgrade defaults preserve to true (keep EPHEMERAL partition) — differs from talosctl default of false
Dry-run defaulttalos_patch_config defaults to dry_run=true; applying requires both dry_run=false and confirm=true
Audit loggingAll mutating tool calls (talos_service_action, talos_reboot, talos_upgrade, talos_rollback, talos_reset, talos_patch_config, talos_apply_config) emit a structured log line to stderr: AUDIT timestamp=<RFC3339> tool=<name> nodes=<list> args=<json> (patch content is redacted)

What Is Not in the Threat Model

  • The LLM itself — prompt injection, hallucinated tool arguments, and LLM provider data retention are outside the scope of this server
  • The MCP client — security of Claude Code, Codex, or other MCP clients is the responsibility of those projects
  • Network path between talos-mcp and Talos nodes — protected by mutual TLS using the credentials in your talosconfig

Least-Privilege Credential Setup

Create a dedicated talosconfig with minimal permissions for use with this server:

Read-only access (recommended for most use cases):

# Generate a reader-only talosconfig
talosctl config new --roles=os:reader talosconfig-readonly

Then set TALOSCONFIG=/path/to/talosconfig-readonly and TALOS_MCP_READ_ONLY=true for maximum restriction. With this setup, the server exposes only read-only tools and the credentials cannot perform any mutating operations even if a tool were somehow bypassed.

Operator access (for service management, reboot, upgrade):

talosctl config new --roles=os:operator talosconfig-operator

This covers all tools except talos_patch_config (which requires os:admin).

Full access (required for config patching):

Use your default talosconfig or generate one with os:admin. Reserve this for setups where config patch capability is explicitly needed.

Verifying Downloads

Checksums (integrity)

Each release includes a talos-mcp_<version>_checksums.txt file with SHA-256 hashes of all archives. Verify the binary after downloading:

# Download archive and checksums
curl -LO https://github.com/Nosmoht/talos-mcp-server/releases/download/v<version>/talos-mcp_<version>_linux_amd64.tar.gz
curl -LO https://github.com/Nosmoht/talos-mcp-server/releases/download/v<version>/talos-mcp_<version>_checksums.txt

# Verify
sha256sum --check --ignore-missing talos-mcp_<version>_checksums.txt

This detects corruption or truncated downloads. It does not protect against a compromised release pipeline.

GitHub Artifact Attestations (SLSA L2 provenance)

Each release includes a GitHub-native build provenance attestation that cryptographically links the binary to the specific commit and workflow run that produced it:

gh attestation verify talos-mcp_<version>_linux_amd64.tar.gz \
  --repo Nosmoht/talos-mcp-server

This requires the GitHub CLI. A passing verification means the artifact was produced by the official release workflow in this repository, not a third-party build.

npm Package Provenance

The npm package is published with provenance attestation:

npm audit signatures

A passing result means the package was published by the official GitHub Actions release workflow via OIDC trusted publishing.

Development

# Build
go build -o talos-mcp ./cmd/talos-mcp

# Test
go test -race ./...

# Lint (requires golangci-lint v2)
golangci-lint run

# Format check
gofmt -l .

License

MIT

Recommend Servers
TraeBuild with Free GPT-4.1 & Claude 3.7. Fully MCP-Ready.
CursorThe AI Code Editor
Playwright McpPlaywright MCP server
MCP AdvisorMCP Advisor & Installation - Use the right MCP server for your needs
Y GuiA web-based graphical interface for AI chat interactions with support for multiple AI models and MCP (Model Context Protocol) servers.
Zhipu Web SearchZhipu Web Search MCP Server is a search engine specifically designed for large models. It integrates four search engines, allowing users to flexibly compare and switch between them. Building upon the web crawling and ranking capabilities of traditional search engines, it enhances intent recognition capabilities, returning results more suitable for large model processing (such as webpage titles, URLs, summaries, site names, site icons, etc.). This helps AI applications achieve "dynamic knowledge acquisition" and "precise scenario adaptation" capabilities.
Tavily Mcp
Baidu Map百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
ChatWiseThe second fastest AI chatbot™
Serper MCP ServerA Serper MCP Server
RedisA Model Context Protocol server that provides access to Redis databases. This server enables LLMs to interact with Redis key-value stores through a set of standardized tools.
AiimagemultistyleA Model Context Protocol (MCP) server for image generation and manipulation using fal.ai's Stable Diffusion model.
MiniMax MCPOfficial MiniMax Model Context Protocol (MCP) server that enables interaction with powerful Text to Speech, image generation and video generation APIs.
WindsurfThe new purpose-built IDE to harness magic
Jina AI MCP ToolsA Model Context Protocol (MCP) server that integrates with Jina AI Search Foundation APIs.
Amap Maps高德地图官方 MCP Server
DeepChatYour AI Partner on Desktop
Howtocook Mcp基于Anduin2017 / HowToCook (程序员在家做饭指南)的mcp server,帮你推荐菜谱、规划膳食,解决“今天吃什么“的世纪难题; Based on Anduin2017/HowToCook (Programmer's Guide to Cooking at Home), MCP Server helps you recommend recipes, plan meals, and solve the century old problem of "what to eat today"
BlenderBlenderMCP connects Blender to Claude AI through the Model Context Protocol (MCP), allowing Claude to directly interact with and control Blender. This integration enables prompt assisted 3D modeling, scene creation, and manipulation.
EdgeOne Pages MCPAn MCP service designed for deploying HTML content to EdgeOne Pages and obtaining an accessible public URL.
Visual Studio Code - Open Source ("Code - OSS")Visual Studio Code