CustoVox MCP servers
custovox.ai — open-source customer-intelligence MCP servers
Free MCP servers that let Claude (or any MCP client) read what customers say: sentiment, emotions, support tickets, reviews, churn signals, photos and calls. They run on your machine with open models.
Built by CustoVox.ai.
MCP servers
Server | What it does | Languages | PyPI |
5-level sentiment of customer text, batch up to 50 | One multilingual model (EN, FR, DE, ES, PT among others) | ||
7 emotions (joy, anger, fear, sadness, surprise, disgust, neutral) | English only | ||
Ticket priority, category, sentiment and escalation risk | Sentiment multilingual; priority and category best in English | ||
Sentiment breakdown and summaries of up to 100 reviews | Sentiment multilingual; summaries best in English | ||
Rule-based churn-risk score 0–100 from a customer's messages | Sentiment multilingual; intent detection best in English | ||
First triage of customer photos (rooms, returns, retail) | English captions | ||
Call analysis: sentiment, contact reasons, churn threats, follow-up actions; demo mode | Simple rules, French and English |
All seven are listed in the official MCP registry under io.github.jeremdecustovox/.
Related MCP server: MCP Sentiment Analysis Server
Library
Package | What it does | PyPI |
Python library behind custovox-calls (not an MCP server): one JSON per call with transcript, agent/customer turns, sentiment curve, contact reasons and key moments |
Quick start
claude mcp add custovox-calls-demo --env CUSTOVOX_CALLS_DEMO=1 -- uvx custovox-calls
# Then ask Claude: "Which demo calls had a churn threat, and what should we do?"Each server's README has its own quick start, tools and limits.
Limits
These servers are meant to sort and prioritize customer feedback for human review. They use open models and simple, readable rules; several parts work best in English (see each README).
More
The hosted platform at custovox.ai covers 5 languages, voice, text and images, with reports and integrations.
MIT License.
Available Tools
4 toolsanalyze_callA
Analyze one customer-service call and save the result in the local database.
source: a local audio file, an http(s) URL to an audio file, a local JSON file of transcript
segments [{start, end, text}], or a demo call id (demo-1 ... demo-6).
started_at: ISO date-time of the call (default: now). language: e.g. "fr" or "en" (audio only).
Returns sentiment, contact reasons, key moments (complaint, churn threat, agent promise) and a
transcript excerpt. Runs locally: audio is transcribed on this machine.Inputs: source (required), call_id, language, started_at.
get_call_insightsA
Summary of the analyzed calls over a period: number of calls, minutes, mean customer sentiment, contact reasons, key moments, churn threats and the most negative calls. since / until: ISO dates, e.g. "2026-11-01" (until is exclusive). Read-only.
Inputs: since, until.
search_feedbackA
Search what was said in the analyzed calls. All words must appear in the same sentence; case and accents are ignored. speaker: "customer" (default), "agent", or null for both. Returns the matching sentences with their call id and time. Read-only.
Inputs: query (required), limit, since, until, +1 more.
suggest_actionsA
Suggest follow-up actions for an analyzed call: manager alert (churn threat or complaint), customer callback (to check an agent promise is kept), CRM ticket (negative sentiment). Each action has a priority, a reason and the sentence that triggered it. Sends nothing.
Inputs: call_id (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
Each tool has a distinct role: analyze_call writes/ingests, get_call_insights aggregates, search_feedback does sentence-level lookup, and suggest_actions derives recommendations for one call. The main soft spot is that suggest_actions implicitly requires a call already analyzed by analyze_call (and a call id), and get_call_insights vs search_feedback both report on past calls, though their granularity is clearly described.
All four names follow a clean verb_noun snake_case pattern: analyze_call, get_call_insights, search_feedback, suggest_actions. The verb choice also signals intent well (analyze=write, get=read aggregate, search=lookup, suggest=derive), with no style mixing.
Four tools is a tight, well-scoped surface for a local call-analytics server, and each tool earns its place. It sits at the low end of the ideal range, with no explicit list/get-calls tool to round it out.
Core flow (analyze → aggregate → search → suggest) is covered, but there is no way to fetch a single stored call's analysis by id, nor to delete/re-analyze a call, which matters for a local database-backed tool. Agents must rely on aggregate or search results to recover call ids.
Maintenance
Related MCP Connectors
One MCP server for all web data. Scrape, search, crawl, and AI research in one install.
Unified 15-tool SEO, SERP, clean Markdown & GEO (LLM citation Share-of-Voice) MCP server.
Free OpenAI-compatible inference with signed provenance receipts and 3 focused MCP tools.
Structured analysis API and remote MCP tool for text, JSON records and numeric series.
Related MCP Servers
- AlicenseBqualityCmaintenanceContent moderation and profanity detection MCP server with 19 tools, 24 language support, leetspeak/Unicode obfuscation detection, context-aware analysis, batch processing, and user tracking for AI-powered content safety.22421 npm69MIT
- AlicenseNot gradedqualityDmaintenanceA robust MCP server for real-time sentiment analysis with AI-powered insights, featuring multi-dimensional emotion detection, batch processing, and a Gradio web interface.3MIT
- FlicenseNot gradedqualityNot gradedmaintenanceEnables aspect-level sentiment mining of customer reviews, producing satisfaction radar breakdowns and structured sentiment analysis through MCP-compatible clients.-
- FlicenseNot gradedqualityNot gradedmaintenanceEnables analysis of acoustic vocal emotion, prosodic valence, and customer frustration from audio inputs through a zero-dependency MCP server.1-