
UniFi Network MCP
io.github.sirkirby由 GitHub 帳號 sirkirby 發布v0.7.8更新於 Oct 24, 2026
Manage UniFi Network devices, clients, firewall, VLANs, VPNs, and more via MCP.
概覽
提供 210 個 UniFi Network 控制器工具的 MCP 伺服器,讓助理檢視並安全地修改裝置、用戶端、防火牆、VLAN 與 VPN。
- 可以做什麼
- 查詢用戶端、裝置、網路、WLAN、統計、流量流與近期事件
- 讀取並稽核防火牆規則、區域、NAT 規則、VLAN 與 VPN 設定
- 以預覽後確認的方式建立、更新、啟停與刪除防火牆、NAT 與 VPN 設定
- 重新命名或修改裝置、封鎖用戶端,並回讀已保存的欄位
- 支援批次呼叫工具與按需載入的工具探索,便於較大型工作流程
- 適用情境
- 適合讓助理回答 UniFi 網路相關問題,或在控制器上執行防火牆、VLAN、VPN 與裝置變更,包括健康檢查與防火牆稽核。若只需從自有軟體透過 REST 或 GraphQL 存取,則不需要使用。
- 執行需求
- 以 stdio 在本機執行,可從 PyPI(uvx 或 pip)或 Docker 安裝。需要能連線到控制器,並提供主機、連接埠與站點,以及管理員憑證 UNIFI_USERNAME 和 UNIFI_PASSWORD;也可選用 UNIFI_API_KEY 做唯讀清單查詢。部分寫入操作需要工作階段憑證。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 UniFi Network MCP,將其新增到工作區。
- 為需要使用其工具的對話啟用該服務。
Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。
其他 MCP 客戶端
參照 儲存庫 中的啟動說明。
README
UniFi Network MCP Server
MCP server exposing 210 UniFi Network Controller tools for AI assistants and other MCP-capable clients. Query clients, devices, firewall rules, VLANs, VPNs, Traffic Flows, stats, and more with safe-by-default permissions and preview-before-confirm for all mutations.
Use this server when an existing AI assistant or other MCP-capable client needs
UniFi Network tools. Software integrating over HTTP, including automation
services, code-execution runtimes, and MCP adapters, should use
unifi-api-server for REST and GraphQL reads, actions, and SSE
events. The calling application must sandbox generated code and enforce its
resource limits.
Install
Claude Code (recommended)
The plugin installs the MCP server, an agent skill for tool discovery, and a guided setup command:
Then run the interactive setup to configure your controller connection:
This walks you through entering your controller host, credentials, and permission preferences. Credentials are given as a file or command provider, never typed into chat. Setup saves the result as the plugin's Claude Code options, which apply in every project; a password or API key entered as an option is stored in the system keychain. Start a new Claude Code session after setup to connect.
Upgrading from a plugin version before 1.0.0, which kept settings in the project's .claude/settings.local.json? Follow the plugin upgrade guide.
Codex
Register the marketplace, then install unifi-network from Codex's /plugins UI:
After installing, ask Codex to use the UniFi Network setup skill. Setup saves a version-pinned MCP server entry with your controller settings in Codex's config.toml and prompts you to restart Codex. After each plugin upgrade, re-run setup with --refresh.
PyPI / Docker
Usage Examples
For the approved breaking MAC argument rename in the next minor release, see the MAC parameter migration guide.
Once connected, just ask your AI agent in natural language:
"Show me all clients on the Guest VLAN with their signal strength and data usage"
"Create a firewall rule that blocks IoT devices from reaching the internet between midnight and 6 AM"
"Which access points have the most client disconnections this week?"
"Audit my firewall policies — are there any redundant or conflicting rules?"
"Rename the device at 192.168.1.45 to 'Living Room TV' and show me its traffic stats"
"What changed on my network in the last 24 hours? Show me new clients and config changes."
"Show me the largest traffic flows from the last hour and summarize who talked to what."
All mutations (firewall rules, device changes, client blocking) use a preview-then-confirm flow — you see exactly what will change before anything is applied.
V2 NAT rules
unifi_list_nat_rules and unifi_get_nat_rule read V2 NAT rules. The create,
update, delete, and toggle tools require Network session credentials and use
V2 NAT rule IDs; port-forward and Integration API IDs are different. Public
writes cover DNAT, SNAT, and MASQUERADE with IPv4, tcp_udp or all, and
NONE or ADDRESS_AND_PORT filters. Creates default to enabled=false.
Created rules are marked manual and non-predefined, and both values are checked
on readback. Existing rules with missing or non-manual origin are not editable
through these tools.
SNAT rules with a translated port require source_filter.port; the controller
validates port correspondence. Existing rules missing that source port must be
corrected in the same update before they can be edited through these tools.
Toggle takes an explicit enabled boolean, so repeating the same request is
a no-op. IPv6, PPPoE, inverted/excluded matching, and other filter variants
are not verified for public writes.
Selector edits on a stored rule with an unverified type are refused, even when
the same request changes its type to a verified one.
The MCP update preview shows a fresh before/after merge. Confirmation reads fresh state again, sends at most one full replacement, then checks persistence; delete checks fresh absence. If the controller reply or readback is uncertain, list rules before retrying. A concurrent controller edit between fetch and PUT can be overwritten because V2 NAT has no conditional replacement. These checks prove stored configuration, not packet-level enforcement. Existing Core NAT methods retain their return shapes; controller mutation exceptions now carry safe fixed guidance instead of raw response text.
For an approved packet test, use the IPv4 DNS redirect recipe. It covers a single client and synthetic destination, including same-subnet return traffic and cleanup.
Configure
Set these variables in the server's process environment. The plugin setup skill already does this for Codex and OpenClaw; in Claude Code the plugin maps its options onto these variables. The server does not automatically load .env or working-directory YAML files; load a trusted env file explicitly in your launcher (Docker env_file: is supported), or select custom YAML with an absolute CONFIG_PATH. See configuration for migration examples.
With no controller host configured, the server still starts in setup mode: it tells the client how to configure it, offers only the read-only unifi_setup_status tool, and makes no controller connections.
Fallback: Existing UNIFI_* variables (e.g., UNIFI_HOST) continue to work. The server checks for UNIFI_NETWORK_* first and falls back to UNIFI_* if the server-specific variable is not set. For single-controller setups, the shared variables are all you need.
API-key inventory and session authentication
Configure UNIFI_NETWORK_API_KEY without a username/password for device,
active-client, network, and WLAN lists. The same tools select the available
read path; no duplicate API-key tools are needed. When valid session credentials
are also configured, the existing session path remains preferred.
On controllers accepting API keys on legacy inventory endpoints, the usual IDs
and fields are preserved. This was verified on Network 10.6.106 with a UniFi OS
proxy; the minimum supported firmware is not established. Other controllers
can use the public Integration inventory fallback when that API is available.
Public inventory has narrower fields and network coverage (for example, WAN/VPN
networks may be absent). Results identify source_api=integration and expose
the public UUID as integration_id, never as a legacy resource ID. Do not pass
these UUIDs to legacy details/update tools. Missing legacy fields are unknown.
API GraphQL projections therefore allow null for unavailable enabled, wired,
and guest flags; clients must handle those nulls.
Historical clients (include_offline=true), full details, legacy mutations,
and the Network websocket still require their supported legacy/session paths.
API-key legacy transport is restricted to verified inventory GET requests;
successful reads do not imply permission to write. Explicit Integration tools
such as DPI lookup and firewall ordering retain their own API-key contracts
(legacy zone-ID resolution also needs a session). Protect still requires session
bootstrap, with an additional API key for public setters; Access retains its
independent API-token and proxy-session paths. Discovery reports local_only,
api_key_only, either, or both; these describe requirements, not proof that
a configured credential has authenticated successfully.
MCP response size
For tool results that already provide structured output, adaptive is the default response mode. It classifies each request by the canonical date-based protocolVersion advertised during MCP initialization, not by the client's product name or application version. Requests advertising MCP 2025-06-18 or later receive concise content plus the full result once in structuredContent; requests advertising an earlier revision (such as 2024-11-05 or 2025-03-26), or whose revision metadata is missing or malformed, keep full compatibility JSON in content. Set UNIFI_NETWORK_MCP_CONTENT_MODE to compat to force the duplicated compatibility form, or to compact to force concise text plus full structured output outside a negotiated request. This server-specific variable overrides UNIFI_MCP_CONTENT_MODE; use compat for any client that consumes the full result only from content, regardless of its advertised revision.
The lazy-loading meta-tools (*_tool_index, *_execute, *_batch, *_batch_status, and lazy-only *_load_tools) remain content-only; they are not the pre-2025-06-18 protocol category described above. For structured inner results, *_execute and *_batch_status expose one normalized JSON payload in content rather than a nested transport pair; content-only execute results remain unchanged. Response modes do not convert these meta-tools to structuredContent.
Network also bounds two large source responses by default: unifi_get_dashboard uses summary=true, and unifi_list_rogue_aps returns a summarized page of at most 100 records. Pass summary=false to request the full selected dashboard or rogue-AP data.
Sensitive response fields
Network tools redact known secret-bearing fields by default before returning data to MCP clients. This includes WLAN passphrases, VPN key material, whole VPN config blobs, SNMP community strings, SNMPv3 passwords, and device-SSH credentials and the management key in raw/detail responses and mutation previews. Disable redaction for a trusted local administration process with UNIFI_NETWORK_REDACT_SENSITIVE_FIELDS=false or the global UNIFI_REDACT_SENSITIVE_FIELDS=false policy flag when raw values are required.
Confirmed Network/WLAN and VPN-state writes are re-read from the controller and report exact persisted_fields, unchanged_fields, dropped_fields, and coerced_fields. Already-satisfied no-op fields appear under unchanged_fields and do not make a failed write partially successful. A response with success: false and mutation_applied: true means the controller accepted the mutation but did not persist it exactly; it may be a partial write, not a rollback. Inspect partial_success and details_after_attempt before retrying or applying a compensating change.
unifi_update_vpn_server_alternate_address updates a WireGuard server's advertised address using server_id, a partial update_data object, and confirm (default false). The accepted fields are alternate_address_enabled (boolean) and alternate_address (hostname or IPv4 address, without a URI, port, or path). Enabling requires a valid address in the merged configuration. Disabling preserves the stored address unless you also supply a replacement. Null and unknown fields are rejected. Server reads expose these same names, with null for absent controller values.
This operation requires Network session credentials and the VPN_SERVERS UPDATE policy gate. MCP previews show fresh before/after values; API action previews show only submitted arguments, with effective-state validation at execution. The confirmed update fetches the full record again, preserves unrelated fields, and verifies the requested fields after writing. Full-record PUTs can still race with concurrent edits. An uncertain transport or readback outcome returns mutation_applied: null; inspect the server before trying again. Previews and results contain only the two public alternate-address fields. Other VPN server types are outside this verified write scope.
Run
Claude Desktop
Add to claude_desktop_config.json:
Agent Skills
When installed via Claude Code, the network plugin ships three agent skills that extend Claude with specialized workflows for network management.
Network Health Check
Trigger: "check network health", "what's down", "run a health check", "network status"
Gathers a full diagnostic snapshot in a single unifi_batch call — system info, network health, device list, and active alarms — then produces a structured health report. Includes reference documents for device state codes, alarm types and severity levels, and health subsystem diagnostics (WAN → LAN → WLAN → VPN priority order).
Firewall Manager
Trigger: "block traffic", "create firewall rule", "set up IoT isolation", "manage content filtering"
Natural-language firewall management with a safe preview-then-confirm workflow. Ships with:
-
Policy templates for common scenarios (
references/policy-templates.yaml): -
Snapshot/diff workflow — Claude saves a timestamped JSON snapshot of all policies, zones, and groups before every mutation, then diffs the after-state against it so unintended changes are caught immediately
-
Reference docs for firewall schema, DPI categories, and full template parameter lists
Firewall Auditor
Trigger: "audit firewall", "review firewall rules", "check for security issues", "score my firewall"
Comprehensive automated audit across 16 security benchmarks in 4 categories, producing a 0–100 score with per-finding remediation guidance. Claude dispatches the MCP tool calls and evaluates each benchmark against references/security-benchmarks.md; a small CLI (scripts/unifi-firewall-score) turns the findings into the canonical, version-stable score so audit history stays comparable.
Score thresholds:
Benchmark categories (4 × 25 points):
- Segmentation (SEG-01–04) — IoT/Guest/Management VLAN isolation, explicit inter-VLAN policies
- Egress Control (EGR-01–03) — Outbound filtering for high-risk VLANs, DNS enforcement, threat intelligence blocks
- Rule Hygiene (HYG-01–05) — Conflicts, redundant/disabled rules, stale references, naming, shadowing
- Topology (TOP-01–04) — Offline devices, firmware currency, VLAN consistency across switch uplinks, orphaned port profiles
Each finding includes the benchmark ID, severity (critical/warning/informational), a plain-language explanation, and — when automatable — the exact MCP tool call to fix it. Audit history is tracked in audit-history.json so score trends are visible over time.
Tool Improvements
Device Classification (unifi_list_devices)
unifi_list_devices now returns a device_category field that correctly classifies every adopted device:
The ap category uses the controller's is_access_point boolean flag as the authoritative signal, not just the device type prefix. This means USP Smart Power strips — which appear as uap-typed devices — are correctly excluded from the AP category.
Enriched Device Fields
Each device record now includes additional fields alongside the existing MAC, name, model, IP, firmware, uptime, and status:
Documentation
- Configuration — Full env var reference, YAML config, controller type detection
- Permissions — Permission system, category defaults, how to enable high-risk tools
- Tool Catalog — All 210 tools organized by category
- Transports — stdio, Streamable HTTP, and SSE setup
- Troubleshooting and support bundles — Reviewed support evidence, connection issues, SSL, missing tools
Development
See the root CONTRIBUTING.md for the full monorepo workflow.
License
來源:apps/network/README.md,提交 fbd083c
工具
0版本歷史
1- v0.7.8最新Sep 16, 2026
評分
已滿足適用檢查項的 50%
該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。
- 近 30 天內可連通25 分
請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。
- 已宣告認證方式10 分
請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。
- 已列出工具15 分
請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。
- 描述必需15 分
- 圖示10 分
- 文件15 分
- 授權條款10 分