接入你的智能体
發布者自己的指令。不是 PlugMCP 端點。
用我們的金鑰與日誌託管一份副本。
npx -y @cognigy/[email protected]Claude Code 沒有本機指令的安裝連結。複製下方指令並在終端機中執行。
claude mcp add cognigy-plugin-engine -- npx -y @cognigy/plugin-engine@1.23.1在終端機中執行。
{ "mcpServers": { "cognigy-plugin-engine": { "command": "npx", "args": [ "-y", "@cognigy/[email protected]" ] } }}claude_desktop_config.json
遠端 URL 還是本機指令？ 兩種安裝方式的差異
來自發布者
Distributed exclusively through each client's native plugin mechanism — a plugin on Claude Code, Claude Desktop, ChatGPT + Codex, and Antigravity, a plugin on Cursor, and an -standard plugin on any conformant host (, , …) — with more clients to come. Each package installs the server engine and ships skills + agents.
MCP server engine for the NiCE Cognigy Plugin (Claude Code, Codex, and more)
A plugin that connects your AI assistant to the Cognigy.AI REST API. Create, test, and improve LLM-based AI Agents through a self-improvement loop — without leaving your client.
Quick links: Installation · Staying up to date
Two steps. Step 1 is the same for every client. Step 2 is where you check what your client still needs from you — for most clients that's nothing, but the answer is in the table, so look your client up rather than assuming.
One installer covers every client. Run it, pick your client(s), enter your Cognigy API base URL (Enter for the trial default) and API key (masked as you type), then restart the client.
macOS / Linux — one line (checks for Node.js, then runs the installer):
bash <(curl -fsSL https://raw.githubusercontent.com/Cognigy/cognigy-plugin/main/install.sh)
Windows (PowerShell) — open as Administrator, then:
irm https://raw.githubusercontent.com/Cognigy/cognigy-plugin/main/install.ps1 | iex
The bootstrap only checks that Node.js 20+ is present (it tells you how to install it if not — it never installs it for you), then runs exactly the command below. Use whichever you prefer; they do the same thing.
Already have Node.js 20+? Skip the bootstrap:
npx -y -p @cognigy/plugin-engine@latest cognigy-setup
The installer wires your credentials everywhere and installs the plugin wherever it can, but some clients need one manual step from you. Find your client below and do what Extra step after the installer says — None means step 1 finished the job. Each guide has the full instructions, Windows notes, how to verify it worked, and troubleshooting.
Cursor is the one client that needs no installer at all: it collects your credentials itself, so you can start at its guide and skip step 1.
1 Claude Code leaves auto-update off for third-party marketplaces — turn it on once under /plugin → Marketplaces → cognigy-plugin.
2 Cursor asks for the credentials itself (Plugins → Configure), so it doesn't need the installer — see its guide.
3 The host owns the plugin version.
Skip the prompts with flags:
npx -y -p @cognigy/plugin-engine@latest cognigy-setup \
--client claude-code --client claude-desktop --client codex \
--client antigravity --client other-hosts \
--api-base-url https://api-trial.cognigy.ai --api-key <key>
--client is repeatable; valid values are claude-code, claude-desktop, codex, antigravity, and other-hosts. Omit it in interactive mode to pick from the menu.
See the Auto-updates column in the table above for what each client does on its own. To pull a new release now instead of waiting for the client to notice:
npx -y -p @cognigy/plugin-engine@latest cognigy-setup update
The installer doubles as a manager:
npx -y -p @cognigy/plugin-engine@latest cognigy-setup status # what's installed, per client, + latest available
npx -y -p @cognigy/plugin-engine@latest cognigy-setup update # pull the latest where it isn't automatic
npx -y -p @cognigy/plugin-engine@latest cognigy-setup uninstall # remove from every client (--client narrows it; --purge also clears ~/.cognigy-plugin)
Beyond the MCP tools, the plugin ships skills and agents that surface the workflows automatically:
/skills) — one per workflow (agent creation, knowledge/RAG, voice gateway, voice go-live checklist, webchat, flow nodes, packages, settings, LLM providers, tools, red-teaming, troubleshooting). Claude loads the matching skill automatically when your request fits. Red-teaming runs in your main session — not as a subagent — so it can agree the scope contract, get approval before applying fixes, and open the finished report for you./agents) — cognigy-agent-builder runs the full build-and-test loop for a new agent, and cognigy-voice-go-live audits a voice agent against the Go-Live Checklist and applies the safe fixes. Each runs in its own context and reports back a summary.The installer collects your Cognigy API base URL and API key and wires them per client: on Claude Code into the system keychain, on Claude Desktop into claude_desktop_config.json (chmod 600). The engine receives them as environment variables. If either is missing for a given launch, the engine falls back to ~/.cognigy-plugin/config.json (chmod 600), which the installer also writes — so credentials resolve from the environment first, then that file. ChatGPT + Codex, Antigravity, and other hosts rely on that file exclusively: their client configs carry no secrets.
A value that arrives as an unexpanded ${...} placeholder counts as missing. Hosts other than Claude Code don't implement userConfig and pass the manifest's ${user_config.cognigy_api_key} through literally; treating that as a real credential would both fail the request and shadow the file fallback. The optional variables below can be set in the MCP server env if you need to override defaults.
Everything this plugin changes is recorded in Admin Center → Audit Events with performedBy.actor = "mcp-plugin", so plugin-made changes are distinguishable from ones people made by hand (which carry no performedBy). Each MCP tool call gets its own taskId and each engine run its own sessionId, so the writes of one tool call group together.
Requires Cognigy.AI 2026.17.0 or newer; older versions ignore the attribution and record the action normally. The "Performed by" column in the UI is gated behind the platform's FEATURE_ENABLE_PLATFORM_AGENT_MFE flag, but the attribution is always recorded and always readable through list_resources { resourceType: "audit_event", actor: ["mcp-plugin"] }.
Set COGNIGY_DISABLE_AUDIT_ATTRIBUTION=1 to opt out; plugin actions are then recorded like any other API action.
If your network requires a proxy, set it in the MCP server env block (or export it before starting your client) and the plugin tunnels every Cognigy request through it:
"env": {
"HTTPS_PROXY": "http://proxy.corp.example:8080",
"NO_PROXY": "localhost,127.0.0.1",
"NODE_EXTRA_CA_CERTS": "/etc/ssl/certs/corporate-root-ca.pem"
}
Lower-case spellings (https_proxy) work too, as do credentials in the URL (http://user:[email protected]:8080). HTTP and HTTPS proxies are supported; SOCKS proxies are not.
The proxy is chosen per request, so NO_PROXY can exclude one host while another still goes through the proxy — a package download link, for example, points at wherever the platform staged the archive rather than at the API host.
If a proxy is configured but cannot be used — a malformed URL, or a SOCKS proxy — requests fail with a configuration error rather than quietly connecting directly, which would send your API key outside the sanctioned path. Exclude the host with NO_PROXY if you genuinely want a direct connection.
A proxy that accepts the connection and then never completes the tunnel would otherwise hang a tool call indefinitely, so connecting and negotiating is bounded at 30 seconds. Raise COGNIGY_PROXY_CONNECT_TIMEOUT_MS if your proxy is simply slow.
If your proxy inspects TLS (Zscaler, Netskope, BlueCoat and similar), it presents its own certificate signed by a corporate root CA that Node does not trust by default. Point NODE_EXTRA_CA_CERTS at that CA file — Node reads it only at startup, so restart your client afterwards. Never set NODE_TLS_REJECT_UNAUTHORIZED=0 instead; that disables certificate verification for every connection the engine makes.
To confirm the engine picked the settings up, set LOG_LEVEL=debug and look for Cognigy API requests route through a proxy in your client's MCP log. GUI clients (Claude Desktop, Antigravity) start the engine with a minimal environment and often do not inherit your shell's proxy variables, so set them in the config file rather than your shell profile.
The engine talks to more than one Cognigy host: the API for everything, the endpoint host for talk_to_agent, and the webchat and static hosts for embed snippets. The other three are derived from COGNIGY_API_BASE_URL by swapping its api segment — api-dev.cognigy.ai gives endpoint-dev.cognigy.ai, cognigy-api-na1.nicecxone.com gives cognigy-endpoint-na1.nicecxone.com, and a local cluster's api.test gives endpoint.test. If your hostnames follow no such pattern, set COGNIGY_ENDPOINT_BASE_URL, COGNIGY_WEBCHAT_BASE_URL and COGNIGY_STATIC_FILES_BASE_URL explicitly; the engine warns at startup (Could not derive the endpoint host) when it had to fall back to the API host.
A cluster that serves a self-signed certificate is rejected with DEPTH_ZERO_SELF_SIGNED_CERT, because Node does not consult the operating-system trust store. Point NODE_EXTRA_CA_CERTS at the certificate (or the local CA that issued it), exactly as for a TLS-inspecting proxy above.
Create a Cognigy AI Agent called "Support Bot" in project <projectId> with a helpful
customer support persona. If the project has no working LLM, first look for another
project with an llm_model that has a connectionId, import that LLM and its connection
via manage_packages, and only fall back to setup_llm if no reusable LLM exists.
Return the endpoint URL only after the project has a confirmed working LLM.
The MCP server should first check whether the target project already has a working LLM. If not, it should prefer reusing the required LLM resource set plus its connection resources from another project via manage_packages, and only use setup_llm as the last resort before calling create_ai_agent.
Talk to my Support Bot at <endpointUrl> and ask "How do I reset my password?".
Then update the job description to make the response more concise and actionable.
Talk to it again and compare the responses.
This triggers the self-improvement loop: talk_to_agent → evaluate → update_ai_agent → talk_to_agent again.
Create a knowledge store in project <projectId>, add the URL
https://docs.example.com/faq as a source, then attach it to my Support Bot
so the agent can search it when answering questions.
The server will call manage_knowledge to create the store and ingest the URL,
then create_tool to attach it as a knowledge search tool on the agent's job
node.
If the project will use Knowledge Search, the server should first configure
Knowledge AI Settings with manage_settings { operation: 'set_knowledge_ai', ... } using model IDs from the same project.
For normal AI-agent knowledge flows, this should happen before
manage_knowledge { operation: 'create_store', ... }.
The model roles are different:
knowledgeSearchModelId is a separate Knowledge AI setting for search
behavior, but the accepted model type is instance-dependent.set_knowledge_ai attempt.llm_model resources and try those same-project IDs before
creating a new model.list_resources with
resourceType: 'llm_model' and useCase: 'knowledgeSearch' so the
candidates match the Settings UI dropdown rather than the unfiltered
project-wide LLM list.knowledgeSearchModelId
unless the user explicitly asks for that and existing same-project candidates
have already failed API validation.List all my Cognigy projects and show which AI Agents exist in each one.
Uses list_resources with resourceType: 'project' and resourceType: 'agent'.
Upload the package at /absolute/path/to/support-bot.zip into project <projectId>,
show me the import preview, then import it using the default selections.
Uses manage_packages with operation: 'upload_and_inspect', then operation: 'import'.
Create a package named "support-bot" from these resource IDs in project <projectId>,
include dependencies, and save the zip to /absolute/path/to/exports/.
Uses manage_packages with operation: 'list_exportable' to discover candidates, then operation: 'export', then operation: 'download' when needed.
Set the voice preview provider for project <projectId> to Microsoft, then create
a Voice Gateway endpoint for flow <flowId> named "Support Voice" and give me the
WebRTC demo URL I can open in the browser.
This uses manage_settings with operation: 'set_voice_preview' to configure speech, then manage_voice_gateway to provision a voiceGateway2 endpoint with a webrtcDemoUrl.
Configure Knowledge AI settings for project <projectId> by setting the Knowledge Search
model and the Content Parser. If I am creating a new project and another project already
has a working knowledge setup, reproduce that exact Knowledge Search model and Content
Parser choice in the new project before trying generic defaults.
This uses manage_settings with operation: 'set_knowledge_ai' to configure
knowledgeSearch and contentParser at the project-settings level. Reusing
settings does not mean copying another project's settings alone; the required
model must exist in the target project first, either by import or by creating it there.
For knowledge workflows, the MCP should import the full required source-project model set
in one pass, reuse shared connections only once, and try the imported same-project model
IDs before falling back to setup_llm.
Detailed workflow guidance (agent creation, knowledge/RAG, voice, webchat, A2A, flow nodes, packages, settings, LLM providers, tools, troubleshooting) ships as skills that load automatically when your request matches, in clients that support them (e.g. Claude Code) — see the Skills column in the client table under Installation. The tools are built to be used with these skills: if the assistant skips steps or gets a Cognigy workflow wrong, check that the skills are loaded (in Claude Code, /skills lists them) and reinstall the plugin if they are missing.
This MCP server transmits requests to the Cognigy.AI API endpoint configured via COGNIGY_API_BASE_URL. No data is collected, stored, or shared by this MCP server itself — all data remains between your AI client and your Cognigy.AI instance.
Full privacy policy: https://www.cognigy.com/privacy-policy
docs/ folderMCP server engine for the NiCE Cognigy Plugin (Claude Code, Codex, and more)
{ "mcpServers": { "cognigy-plugin-engine": { "command": "npx", "args": [ "-y", "@cognigy/[email protected]" ] } }}工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式，而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
@cognigy/plugin Engine 透過 stdio 在本機執行。
上次檢查時間：2026-11-15。 MCP server engine for the NiCE Cognigy Plugin (Claude Code, Codex, and more)