Dev builds 132-tool MCP server for Claude to manage Cognigy.AI agents
Tsvetan Gerginov built `cognigy-ai-mcp-management-server`, an open-source, MIT-licensed MCP server with 132 tools that gives AI assistants like Claude programmatic access to the Cognigy.AI Management API, replacing manual UI workflows.
Score breakdown
The server replaces manual Cognigy.AI UI workflows with AI-assistant-driven automation while introducing `dryRun`-by-default and secret-redaction patterns as a concrete model for safely wrapping large enterprise APIs with write access into LLM tooling.
- 01132 tools cover the full Cognigy.AI management surface: flows, NLU, snapshots, RAG stores, analytics, audit logs, and more.
- 02Every mutating tool defaults to `dryRun: true` to prevent accidental writes to production agents.
- 03API secrets are stored only in environment variables and memory; connection secrets are redacted before reaching the model.
Tsvetan Gerginov built `cognigy-ai-mcp-management-server` to give AI assistants like Claude, Claude Code, and Cursor programmatic access to the Cognigy.AI Management API — replacing the manual, click-heavy UI workflows he encountered while building enterprise voice bots and multilingual flows at Deloitte. The server is written in TypeScript, MIT-licensed, and published on npm. Its 132 tools span flows and nodes (full CRUD plus search and AI output generation), intents and NLU training, playbooks and regression testing, snapshots and packages, Knowledge AI/RAG stores (21 tools alone), LLM provider configuration, connections, extensions, contact profiles with GDPR export, analytics, and audit logs.
Because the server can mutate production conversational AI agents, Gerginov built in several safety mechanisms.
Because the server can mutate production conversational AI agents, Gerginov built in several safety mechanisms. Every mutating tool defaults to `dryRun: true`, requiring explicit intent to execute a destructive action. API keys are stored only in environment variables and memory, and connection secrets returned by the Cognigy API are automatically redacted before reaching the model. All tool inputs are validated with Zod schemas, and long-running operations like NLU training and snapshot creation use async polling rather than returning a job ID, preventing multi-step agent workflows from silently desyncing.
The repo ships with a Prism mock server generated from Cognigy's OpenAPI spec, enabling local development without a Cognigy account, backed by 49 tests. TypeScript types are also generated from the OpenAPI spec so that API changes surface as compile-time errors rather than runtime failures. Gerginov notes that Cognigy's parent company NiCE shipped an official MCP server mid-project, but he continued because his implementation offers an independent, MIT-licensed, mock-testable artifact with `dryRun`-by-default semantics. The project is not affiliated with or endorsed by Cognigy or NiCE.
Key facts
- 01132 tools cover the full Cognigy.AI management surface: flows, NLU, snapshots, RAG stores, analytics, audit logs, and more.
- 02Every mutating tool defaults to `dryRun: true` to prevent accidental writes to production agents.
- 03API secrets are stored only in environment variables and memory; connection secrets are redacted before reaching the model.
- 04All tool inputs are validated with Zod schemas before hitting the API.
- 05Long-running tasks like NLU training use async polling instead of returning a job ID.
- 06The repo ships with a Prism mock server and 49 tests generated from Cognigy's OpenAPI spec.
- 07The Cognigy official REST client is declared as a peer dependency (not bundled) due to its proprietary license; the MCP server itself is MIT-licensed.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 13, 2026 · 08:58 UTC. How this works →