MCP-TrueNAS
An MCP server that lets Claude monitor and manage a TrueNAS box in plain English. It is deliberately read-only, providing tools to check alerts, storage pools, snapshots, services, and more.
README
MCP-TrueNAS
An MCP (Model Context Protocol) server that lets Claude monitor and manage a TrueNAS box in plain English: "any alerts on the NAS?", "how full is tank?", "list snapshots of appdata".
Read-only by default — the server observes and never mutates unless you opt
in. Safe, reversible writes are enabled with TRUENAS_ENABLE_WRITE=1;
destructive operations are separately gated and human-confirmed. See
Write support & safety.
How it works
┌───────────────── your Windows PC ─────────────────┐
│ │
│ Claude Code ◄── stdio (MCP JSON-RPC) ──► this │ ┌─────────┐
│ (you chat here) server ─┼──────► │ TrueNAS │
│ │ HTTPS/ │ (LAN) │
└─────────────────────────────────────────────────────┘ WSS └─────────┘
- This server runs on your PC, not on the NAS. Claude Code launches it
automatically as a background child process when a session starts (that's
what
.mcp.jsonconfigures) and talks to it over stdin/stdout. - The server talks to TrueNAS over its network API, authenticating with an API key you generate in the TrueNAS UI.
- stdout is sacred: it carries the MCP protocol. All logging goes to
stderr (
console.error), neverconsole.log.
Which TrueNAS API? Both.
TrueNAS changed APIs across versions, so the client auto-detects at runtime:
| Your TrueNAS | API used | How |
|---|---|---|
| 25.04 "Fangtooth" and newer (incl. 26+) | WebSocket JSON-RPC 2.0 at wss://HOST/api/current |
GET /api/versions answers with a JSON array → WebSocket mode |
| SCALE ≤ 24.10, CORE 13.x | REST v2.0 at https://HOST/api/v2.0 |
/api/versions doesn't exist → REST mode |
Version quirks handled for you: the snapshot API rename in 25.10
(zfs.snapshot.query → pool.snapshot.query), {"$date": …} timestamp
objects, and the login-method change coming in TrueNAS 27
(TRUENAS_USERNAME covers it).
File tour
| File | Role |
|---|---|
| src/index.ts | Bootstrap: loads config from .env, creates the McpServer, connects the stdio transport. Small on purpose. |
| src/truenas-client.ts | Everything TrueNAS: API detection, WebSocket JSON-RPC with reconnect + login, REST fallback, and translation of raw network errors into actionable messages. |
| src/tools.ts | The MCP surface — every registerTool() call. To add a capability, you edit this file (and maybe add one method to the client). |
| .mcp.json | Tells Claude Code how to launch this server (project-scoped registration). |
| .env | Your secrets. Gitignored. .env.example is the committable template. |
The tools
All read-only (readOnlyHint: true), all paginate or filter where data can be
large. Each accepts response_format: markdown (default, compact summary) or
json (full structured data). Every tool also declares an outputSchema, so
the structured result is machine-validated by the MCP client.
System & storage
truenas_connection_status— diagnose config/reachability/auth; start here when something failstruenas_get_system_info— version, hostname, uptime, CPU, RAM, loadtruenas_check_updates— whether a base-OS update is available, plus reboot-required statetruenas_list_pools— pool health, capacity, usage %, fragmentation, scrub activitytruenas_list_datasets— space used/available, quotas, compression (filter by pool, paginated)truenas_list_disks— model, serial, size, pool membership, optional temperaturestruenas_list_snapshots— snapshots with creation time and space (filter by dataset, paginated)
Health & activity
truenas_list_alerts— active alerts, filterable by severitytruenas_list_services— SMB/NFS/SSH/… state and boot settingtruenas_list_jobs— background jobs: scrubs, replications, failures
Apps, sharing, backups & VMs — require TrueNAS 25.04+ (WebSocket API)
truenas_list_apps— installed apps and whether an app/image update is available, plus Docker statustruenas_list_shares— SMB shares, NFS exports, and iSCSI targets in one calltruenas_list_network— IP addresses, default routes, DNS, and per-interface link statetruenas_list_replication_tasks— configured ZFS replication (backup) tasks and last-run statetruenas_list_cloudsync_tasks— cloud backup tasks (S3/Drive/B2/…); credentials never showntruenas_list_snapshot_tasks— periodic (automatic) snapshot policy and retentiontruenas_list_scrub_tasks— scheduled pool scrub taskstruenas_list_vms— virtual machines, run state, and resource allocation
Writes (opt-in — set TRUENAS_ENABLE_WRITE=1) — reversible, non-destructive; never marked read-only (so a client won't auto-run them); every call is audit-logged to stderr:
truenas_create_snapshot— take a ZFS snapshot of a dataset (optionally recursive)truenas_update_dataset— change dataset properties (comment, compression, readonly, atime, sync, quota)truenas_control_service— start / stop / restart / reload a service (smb, nfs, ssh, …)truenas_set_service_boot— enable/disable a service starting on boottruenas_manage_app— start / stop / redeploy / upgrade / rollback an installed apptruenas_manage_vm— start / stop / restart / suspend / resume a VMtruenas_run_scrub— start a manual pool scrubtruenas_create_dataset/truenas_rename_dataset— create a dataset or zvol; rename a datasettruenas_create_smb_share/truenas_update_smb_share— SMB sharestruenas_create_nfs_share/truenas_update_nfs_share— NFS exportstruenas_create_user/truenas_update_user/truenas_set_user_password— local users (passwords never logged)truenas_create_group/truenas_update_group— local groupstruenas_create_snapshot_task/truenas_update_snapshot_task— automatic snapshot schedulestruenas_create_scrub_task/truenas_update_scrub_task— scheduled scrubs
Destructive (opt-in — set BOTH TRUENAS_ENABLE_WRITE=1 and TRUENAS_ENABLE_DESTRUCTIVE=1) — irreversible; destructiveHint:true; each requires a human elicitation confirmation and refuses if the client can't be prompted:
truenas_delete_snapshot/truenas_delete_dataset— delete a snapshot / dataset (destroys data)truenas_delete_smb_share/truenas_delete_nfs_share— remove a share (underlying data kept)truenas_delete_app/truenas_delete_vm— delete an app / VM (optionally its data too)truenas_delete_user/truenas_delete_group— delete a user / grouptruenas_rollback_snapshot— revert a dataset to a snapshot (discards newer data)truenas_lock_dataset— lock an encrypted datasettruenas_pool_detach_disk/truenas_pool_offline_disk/truenas_pool_remove_vdev— pool device operationstruenas_upgrade_pool— apply ZFS feature flags (irreversible)truenas_reboot_system/truenas_shutdown_system— reboot / power off the NAStruenas_apply_update— apply the pending OS update and reboot
Write support & safety
The server is read-only until you opt in, in tiers:
| Tier | Enable with | Gate |
|---|---|---|
| Read (default) | always on | none |
| Safe write (reversible) | TRUENAS_ENABLE_WRITE=1 |
not registered unless enabled; audit-logged; never marked read-only |
| Destructive (delete / rollback / reboot) | TRUENAS_ENABLE_DESTRUCTIVE=1 |
the above plus a human elicitation confirmation, and fail-closed if the client can't prompt |
Status: the full read/write surface is implemented — 18 read, 22 safe-write, 17 destructive tools. Destructive tools appear only when BOTH flags are set, and each requires a human elicitation confirmation (fail-closed if the client can't be prompted).
Being honest about enforcement: MCP has no protocol-level way to force human
confirmation, so the guarantees that actually hold are the ones a client cannot
bypass — destructive tools are absent from the tool list unless the operator
sets the flag, and they refuse to run when a human can't be prompted.
Elicitation prompts and Claude Code ask/deny permission rules add
defense-in-depth but are bypassable (--dangerously-skip-permissions,
auto-accepting clients). The most dangerous operations — disk.wipe,
pool.create (formats disks), pool.export with destroy — are deliberately
not implemented; do those in the TrueNAS UI.
Testing writes safely: set TRUENAS_TEST_DATASET=tank/mcp-test to restrict
every write to that dataset and its children, so a mistake can't reach real
data — or point the server at a throwaway TrueNAS VM for write development.
Setup
1. Create an API key on TrueNAS
In the TrueNAS web UI: user icon (top right) → My API Keys → Add.
Copy the whole <id>-<secret> string — it is shown exactly once.
An API key inherits the privileges of the user it's linked to. For a
belt-and-braces read-only setup, create a dedicated user whose group has only
the READONLY_ADMIN privilege (Credentials → Groups/Privileges) and link
the key to that user — then even a leaked key can't change anything.
API keys are password-equivalent and bypass 2FA. TrueNAS auto-revokes keys that ever travel over plain HTTP — always use
https://.
2. Configure
copy .env.example .env
Edit .env: set TRUENAS_URL and TRUENAS_API_KEY; set
TRUENAS_SKIP_TLS_VERIFY=1 if the NAS uses its default self-signed
certificate.
3. Build
npm install
npm run build
4. Register with Claude Code
This repo ships a project-scoped .mcp.json that launches the
server via a relative path (dist/index.js). Once you've built it, Claude
Code asks to approve the server the first time you open a session in this
folder, then launches it automatically thereafter.
To use it from any folder, register it user-wide with the absolute path to your clone:
claude mcp add --scope user --transport stdio truenas -- node /absolute/path/to/MCP-TrueNAS/dist/index.js
Check it connected with claude mcp list (or /mcp inside a session).
5. Use it
Just ask Claude things like:
- "How healthy are my pools?"
- "Anything above WARNING in the NAS alerts?"
- "Which disks are hottest right now?"
- "Are any jobs failing? Show the errors."
Testing without Claude
MCP Inspector is a debugging client. Two flavors:
npm run inspect
opens a web UI (browse tools, call them, see raw protocol traffic), or the scriptable CLI:
npx @modelcontextprotocol/inspector --cli node dist/index.js --method tools/call --tool-name truenas_list_pools
Extending: anatomy of a tool
Every tool in src/tools.ts follows one pattern:
server.registerTool(
"truenas_list_pools", // snake_case, service-prefixed
{
title: "List Storage Pools",
description: "…written FOR the model: what it returns, when to use it…",
inputSchema: z.object({ // zod validates before your code runs
pool: z.string().optional().describe("shown to the model too"),
}),
annotations: { readOnlyHint: true, openWorldHint: false },
},
async ({ pool }) => {
try {
const data = await getClient().pools(); // client hides WS-vs-REST
return respond(structured, format, () => markdown);
} catch (error) {
return errorResult(error); // actionable, never a crash
}
}
);
To add a write tool later (start a scrub, take a snapshot): add a client
method for the API call, register the tool with honest annotations
(readOnlyHint: false, destructiveHint as appropriate) — Claude Code will
then treat it with matching caution. Rebuild (npm run build) and restart
the session; Claude picks up the new tool automatically.
Troubleshooting
| Symptom | Likely cause / fix |
|---|---|
| Any tool errors | Ask Claude to run truenas_connection_status — it diagnoses config, reachability, API mode, and auth in one shot |
| "TLS certificate verification failed" | Self-signed cert on the NAS → TRUENAS_SKIP_TLS_VERIFY=1 in .env |
| "Cannot reach …" | Wrong TRUENAS_URL, NAS off, or firewall |
| "TrueNAS rejected the API key" | Key mistyped/revoked — regenerate; copy the whole <id>-<secret> string |
| Key suddenly stopped working | Was it ever sent over http://? TrueNAS auto-revokes such keys — make a new one, use https:// |
| Tools missing in Claude | .mcp.json not approved yet, or server not rebuilt after edits — claude mcp list shows connection state |
Security notes
- The API key lives only in
.env(gitignored) or the MCP client's env config — never in source, never in chat. - v1 tools are read-only by design; the key's user can enforce the same
server-side via
READONLY_ADMIN. - Plain-HTTP URLs are refused unless
TRUENAS_ALLOW_HTTP=1, because TrueNAS revokes keys observed on unencrypted transport. - TLS verification is on unless you explicitly disable it for a self-signed cert.
- Some TrueNAS read APIs return secrets (cloud-sync OAuth tokens, a VM's VNC
password). The tools that touch those —
truenas_list_cloudsync_tasksandtruenas_list_vms— deliberately project only non-sensitive fields, so no tokens, secrets, or passwords are ever surfaced to the model.
License
MIT © PainInTheNic
推荐服务器
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。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。