v0.1.0

0.1.0 — MCP SDK v2, zod 4, protocol 2026-07-28

The platform release: same 67 tools, same arguments and outputs, new foundation.

Changed

• MCP TypeScript SDK v2 (@modelcontextprotocol/server + @modelcontextprotocol/node, protocol revision 2026-07-28). Clients on older protocol revisions negotiate down automatically — verified with a v1 SDK client: tools, structured output, prompts, resources, and live subscriptions all work.
• zod 4. Every tool input/output schema is now an explicit z.object().
• Behavioral deltas inherited from the SDK (decision D34): listChanged: true is advertised for tools/resources/prompts; the stdio read buffer is capped at 10 MB and non-JSON stdout lines are skipped; HTTP POSTs without Content-Type: application/json answer 415.
• Node floor stays >=20 (D35).

Added

• mcpName in package.json and a server.json for the official MCP Registry (io.github.mswdev/better-unraid-mcp).

Attached

• better-unraid-mcp-0.1.0.mcpb — Claude Desktop one-click bundle (download, open, enter URL + API key).

Full Changelog: v0.0.15...v0.1.0

v0.0.15

0.0.15 — metric history

Added

• Opt-in metrics recorder — set MCP_METRICS_HISTORY=true and the server subscribes to the CPU, memory, and network WebSocket feeds at start and keeps 24 hours of 30-second min/avg/max buckets per topic in memory (2880 slots each; nothing is written to disk; cleared on restart).
• metrics_history (read-only) — topic cpu | memory | network, window_minutes (default 60, max 1440). Concise output folds the window into at most 24 rows; detailed output returns every point with ISO times; missing buckets are counted as gaps, never interpolated. When recording is off the tool explains exactly how to turn it on.
• unraid://live/history resource — the last hour of every recorded topic (or the enable hint).

Internals

• The three metric subscription documents now live in one place (metric-subscriptions.ts) shared by the live resource and the recorder.

Docs

• README configuration row, resources bullet, tool row (67 tools); .env.example; decision log D32–D33.

Full Changelog: v0.0.14...v0.0.15

v0.0.14

0.0.14 — host configuration over SSH

Seven new tools for what the Unraid GraphQL API cannot do, each validated live (read-only) against a real Unraid 7.3.2 server and against the unraid/webgui source, and each verifying its own result by reading it back. 66 tools total.

Services

• service_list — running / stopped / not-installed state of samba, nfs, sshd, docker, libvirt, tailscale via /etc/rc.d/rc.* status (parsed from the status sentence — the scripts exit 0 either way).
• service_action (tier-2) — restart / start / stop with a status read-back. stop also needs allow_stop: true (it cuts clients off their shares); stopping docker or libvirt is refused outright — use array_action or the web UI.

Unassigned devices

• unassigned_list — disks outside the array, pools, and flash (lsblk cross-referenced with emhttpd's disks.ini; pseudo-disks such as zram filtered), with partitions, filesystems, and mount points.
• unassigned_action (tier-2) — mount / unmount a partition through the Unassigned Devices plugin's rc.unassigned, verified with lsblk. Refuses partitions of assigned disks and refuses entirely when the plugin is missing (no raw mount fallback).

Shares

• share_create, share_edit, share_delete (tier-2) — go through emhttpd exactly as the web UI does (the ShareEdit and SecuritySMB forms, via emcmd), expose a validated settings subset (comment, allocator, cache mode/pool, SMB export/security), pass every other setting through unchanged, validate names like the UI (reserved, disk, and pool names refused), and verify each write by re-reading /boot/config/shares/<name>.cfg — any mismatch is reported as "not verified". share_delete only removes the configuration and refuses while the share still holds data.

Docs

• README "Host configuration (SSH)" section; decision log D29–D31; the Phase 3 plan carries the full live-validation record.

Full Changelog: v0.0.13...v0.0.14

v0.0.13

0.0.13 — flash backup

Added

• flash_backup (tier-2, SSH) — archives the Unraid flash device into a user share: /boot/config by default (settings, plugins, keys), or the whole flash with full: true. Writes /mnt/user/<share>/flash-backups/flash-backup-<scope>-<UTC timestamp>.tar.gz, verifies it with tar -tzf, reports size and entry count, and with keep: N prunes older archives while listing exactly what it removed. Progress heartbeats cover the multi-GB full case. The archive contains your license key, API keys, and passwords — it stays on your array; the tool never copies it off the server. This replaces the Unraid API's flash-backup mutation, which is an unimplemented stub upstream.

Decisions

• VM snapshot revert/delete remain deliberately unshipped (D27): the live-validation server has VMs disabled, and the web UI flow depends on webgui-private snapshot metadata, domain-XML rewrites, backing-chain unlinking, and NVRAM handling that cannot be reproduced chain-safely through virsh alone. vm_snapshot_create / vm_snapshot_list still point at the UI for those.

Full Changelog: v0.0.12...v0.0.13

v0.0.12

0.0.12 — DX quick wins

Added

• npx better-unraid-mcp doctor — runs the connection-doctor checks from the command line (GraphQL reachability and versions, schema skew, SSH, rate limit, read-only mode) and exits non-zero when a check fails. README quick start step 3.
• Schema skew check — connection_doctor and the CLI compare the server's API version with the version the vendored schema was fetched from (major.minor, warn-only). npm run schema:update now records that version in schema/schema-version.json.
• Local validation for raw GraphQL — graphql_query and graphql_mutation validate documents against the vendored schema before sending anything (typos get did-you-mean hints); dry_run: true validates only and needs no confirm.

Changed

• Vendored schema refreshed to unraid/api 4.37.4. Consequences:
  • array_disk_action no longer offers remove — upstream retired removeDiskFromArray (unraid/api#2068); a legacy remove call gets an explicit refusal pointing at the web UI.
  • docker_container_action restart uses the API's native restart mutation instead of the composed stop-then-start.
• notification_archive says plainly that the batch API swallows per-id failures (verify with notification_list).

Docs

• Decision log D22–D26; the acceptance-test prompt's archiveAll selection corrected to a schema-valid one.

Full Changelog: v0.0.11...v0.0.12

v0.0.11

0.0.11 — validated on real hardware

The first release exercised end-to-end against a real Unraid box (Unraid 7.3.2, API 4.37.4, 29-disk array, 58 containers, Intel iGPU, NUT-managed UPS). The acceptance run and its findings are in docs/superpowers/specs/roadmap-acceptance-report-2026-09-17.md.

Fixed

• system_health and unraid://health no longer fail when apcaccess has no data. UPS is now read in its own query; when the Unraid API cannot read apcupsd (NUT-managed or absent UPS) the rollup reports ups: UPS data unavailable … instead of erroring out.
• ups_status returns an honest "No live UPS data" result (with upsDetected: false in detailed mode) for the apcaccess-empty case, and explains that NUT-managed UPSes are invisible to the Unraid API.
• disk_list filters the upstream partition-prefix bleed: on servers with more than 26 disks the API attached sdaa1…sdae1 (including the flash drive) to /dev/sda.
• gpu_metrics (Intel) now returns a one-line summary — actual frequency, rc6 idle share, per-engine busy %, GPU clients — parsed from the intel_gpu_top -J stream (newline-separated, truncated by timeout), instead of dumping raw JSON.

Docs

• Acceptance report, Phase 0 plan, decision log D18–D21.

Full Changelog: v0.0.10...v0.0.11

v0.0.10

0.0.10 — Full live coverage + review hardening

New features

• unraid://live/metrics now carries everything — CPU, memory, network, and temperature subscriptions merged into one resource with per-topic ages.
• Three new live resources — unraid://live/ups (UPS telemetry), unraid://live/array (array state/capacity), unraid://live/notifications (newest notification as it arrives).

Fixes (from a full-codebase review of the 0.0.4–0.0.9 work — 14 verified findings)

• Live feed subscriptions now stop when their server/session closes (was: a permanent WebSocket subscription leak per recycled HTTP session); errored subscriptions self-heal on the next subscribe; the docker-stats aggregate evicts containers not seen for 30 s (was: stopped containers reported as fresh "live" data).
• The snapshot cache is cleared after any mutation (was: verification reads within 5 s could serve pre-mutation state).
• system_power now reports SSH auth/DNS failures as real errors (was: a false "the reboot command was issued" success); a remote command that dies without an exit status maps to exit −1, never 0 (was: signal-killed destructive commands reported success); timed-out commands close their SSH channel.
• HTTP session mode: failed initializations no longer leak server/transport pairs, and sessions with an open SSE stream are exempt from the idle sweeper.
• user_script_run rejects ./.. names; system_health handles phantom UPS devices and multi-token apcupsd statuses; the truncation envelope re-shrinks until it fits the output budget; SSH channel-semaphore race fixed.

Changes

• .env.example caught up with all of the roadmap's environment variables.

v0.0.9

0.0.9 — Live telemetry + launch kit

The roadmap's final planned release: real-time data over WebSocket, surfaced the MCP-native way.

New features

• Subscription client — a graphql-ws client (graphql-transport-ws subprotocol, authenticated via the connection_init payload) with lazy connect and bounded auto-retry, feeding an age-stamped live sample store.
• Subscribable MCP resources — unraid://live/parity, unraid://live/docker-stats, unraid://live/metrics, and the templated unraid://logs/{path} log follow. Clients that issue resources/subscribe get notifications/resources/updated on every sample; clients that don't get an honest {"waiting": true} envelope on read.
• docker_stats live preference — serves a fresh subscription-fed sample when one is ≤ 10 s old (labelled source: live subscription, age Xms), falling back to the SSH docker stats snapshot (labelled source: ssh).
• Launch kit — README overhaul (58-tool feature tour, cautious read-only quickstart, per-client install snippets) and a draft Unraid-forum announcement at docs/launch/forum-post.md.

Changes

• New runtime dependencies: graphql-ws, ws.
• The roadmap deliverables shipped: decision log (docs/superpowers/specs/roadmap-decisions.md) and acceptance-test prompt (docs/superpowers/specs/roadmap-acceptance-test.md).

v0.0.8

0.0.8 — Coverage: SSH

Twelve SSH-backed tools that reach past the API. 58 tools. Every tool probes command availability and reports absence clearly; all interpolated values are shell-quoted.

New features

• ZFS suite — zfs_status (pool health + ARC usage from /proc/spl/kstat/zfs/arcstats), zfs_dataset_list, zfs_snapshot_list, and two-flag-gated zfs_snapshot_action (create / destroy / rollback — destroy can only ever target a dataset@snapshot, and rollback deliberately refuses non-latest snapshots).
• gpu_metrics — full per-GPU metrics via nvidia-smi; Intel detection with a bounded intel_gpu_top sample; clear absence report otherwise.
• process_list — top host processes by CPU or memory.
• disk_spin — spin array slots through emhttpd (emcmd, keeping the UI's spin state consistent) or unassigned /dev/sdX via sdspin; NVMe rejected.
• User Scripts — user_script_list, and two-flag-gated user_script_run replicating the plugin's noexec-flash runner (copy to tmp, strip CRs, run via bash, preserve the exit code).
• VM snapshots — vm_snapshot_list and two-flag-gated vm_snapshot_create, using EXTERNAL disk snapshots (virsh snapshot-create-as --atomic --disk-only, the same flow Unraid 7 uses). Revert/delete deliberately stay in the Unraid UI for now (external-chain corruption risk).
• disk_smart_report — the full smartctl attribute report for one disk, tolerating smartctl's informational exit-code bitmask.

v0.0.7

0.0.7 — Coverage: GraphQL

Four new capability areas from the Unraid GraphQL API. 46 tools.

New features

• array_disk_action — add/remove a disk to/from the protected array, mount/unmount an array disk, or clear disk statistics. Array-state preconditions are checked first (membership changes require a STOPPED array; live-disk actions require STARTED), and the tool is two-flag gated — a wrong disk id here risks data loss.
• API key management — apikey_list (metadata only; key values are never selected) and two-flag-gated apikey_manage (create / update / add_role / remove_role / delete). The key value is disclosed exactly once, at creation.
• notification_unread — move a notification back to the unread bucket (reversible, ungated).
• plugin_install_plg — install a native Unraid OS plugin from a .plg URL (two-flag gated; a .plg runs arbitrary code as root).

Changes

• The spec's rclone-remotes/flash-backup item ships nothing, deliberately: upstream's initiateFlashBackup is a Not-implemented stub and rclone is disabled in production builds (decision D9 in the decision log).