Plugin Engine

What @cognigy/plugin-engine does

The @cognigy/plugin-engine MCP server connects an AI assistant to the Cognigy.AI REST API and exposes workflows for building and operating conversational agents. It belongs in the Conversational AI category and covers more than basic agent creation: projects can be configured with LLM resources, knowledge stores, tools, webchat endpoints, voice endpoints, and A2A server endpoints.

The main agent workflow can create an agent together with a flow, an AI Agent job node, and a REST endpoint. After creation, an assistant can change the agent persona, guardrails, job role, procedures, selected LLM, and temperature. The talk_to_agent tool sends test messages through Endpoint Test Mode, which the supplied documentation says is not billable.

Other tools provide resource discovery and administration. They can list Cognigy resources, retrieve one resource in detail, delete resources, manage flow nodes, create or update tools, and create or configure deployment endpoints. Snapshot support provides a way to save and restore project changes.

How it works

The @cognigy/plugin-engine MCP server packages the Cognigy API operations as MCP tools. An AI client invokes a tool with structured input, and the engine validates requests before calling Cognigy. The implementation includes Zod input validation, rate limiting, and RFC 7807-style error responses.

Knowledge workflows can create stores and add URL, text, or file sources, then list chunks for retrieval-augmented generation. Tool management supports HTTP, knowledge, email, MCP, and A2A tool types. Voice workflows can create Voice Gateway endpoints with WebRTC, while webchat workflows target Webchat v3 endpoints for website deployment. A flow can also be exposed as an A2A endpoint for discovery and calls from other agents.

The package includes skills and agents in clients that support them. These provide workflow guidance for activities such as agent building, knowledge configuration, voice readiness checks, webchat, packages, settings, tools, and troubleshooting. It also bundles the official Cognigy documentation MCP server for documentation lookup.

Setup and configuration

Installation is handled through the cognigy-setup installer. Node.js 20 or newer is required. The installer accepts a Cognigy API base URL and API key, then installs or configures the plugin for the selected client. A noninteractive setup can pass client and credential values as command-line flags.

The @cognigy/plugin-engine MCP server requires COGNIGY_API_BASE_URL and COGNIGY_API_KEY when launched with environment configuration. The installer can instead store credentials in ~/.cognigy-plugin/config.json; environment values take precedence, and the file is used as a fallback when a required value is unavailable. LOG_LEVEL, RATE_LIMIT_MAX_REQUESTS, and RATE_LIMIT_WINDOW_MS are optional overrides.

The README documents support for Claude Code, Claude Desktop, ChatGPT with Codex, Antigravity, Cursor, and other Agent Plugins-compatible hosts such as Kiro and VS Code with Copilot. Cursor has its own setup path and asks for credentials in its plugin configuration.

Limitations and notes

The supplied material does not describe a standalone Cognigy account or project provisioning process, so users need access to a Cognigy.AI environment and valid API credentials. Deleting a resource is permanent. Endpoint Test Mode avoids billing for the documented agent conversation test, but that statement does not apply broadly to every Cognigy operation.

Client behavior differs. Some clients install the plugin automatically, while others require an in-app or host-level installation step. Auto-update behavior also depends on the client. The engine falls back when credentials are absent from a supported environment configuration, but an unexpanded ${...} placeholder is treated as missing rather than as a usable key.