Open-source Voice of Customer research agent for product teams.
Ramp built an internal VoC agent that replaces 8 days of research with 8 minutes. This is the open-source version anyone can use with Claude Code.
Connect your customer conversations (Gong, Granola, Jiminy, or plain CSV/JSON), run a slash command, and get structured, evidence-based research output — not summaries.
# 1. Clone
git clone https://github.com/yourusername/vox.git
# 2. Add to Claude Code as a project
cd vox
# 3. Add your data (pick one)
# a) Connect an MCP server (see adapters/)
# b) Drop CSV/JSON files in data/
# c) Use the sample data as-is
# 4. Run a research query
/voc Why are enterprise customers churning?
That's it. No install, no build, no config file. Just Claude Code and your data.
/voc What's the #1 unmet need for teams with 50+ users?
Returns a structured research brief with findings, evidence, confidence levels, contradictions, and follow-up questions. Not a summary — a proper research output.
/persona enterprise product leaders
Returns a persona card grounded in real data: goals, frustrations, behaviors, defining quotes, churn triggers. Every attribute traced to source.
/themes customer calls from Q4
Returns theme clusters with frequency, severity, trend direction, segment analysis, and representative quotes. Uses Braun & Clarke methodology — themes are patterns with a point of view, not topic labels.
/painpoints onboarding experience
Returns scored pain points across four dimensions: frequency, intensity, breadth, and workaround existence. Ranked and classified (Critical / High / Medium / Low).
/jtbd why customers switch from spreadsheets
Returns job statements (functional, emotional, social), forces of progress, opportunity scoring, and competing solutions. Based on Ulwick's ODI and Moesta's Switch methodology.
/opportunity-map reduce enterprise churn to 5%
Returns a Teresa Torres-style opportunity tree: outcome > opportunities > solution hypotheses > assumptions to test. Every opportunity grounded in customer evidence.
| Source | Method | Setup |
|---|---|---|
| Gong | MCP server | adapters/gong.md |
| Granola | MCP server | adapters/granola.md |
| Jiminy | MCP server | adapters/jiminy.md |
| CSV/JSON | Local files | adapters/csv.md |
| Custom | MCP or files | adapters/custom.md |
vox works best with multiple data sources. Cross-referencing findings across sales calls, support tickets, and user interviews produces stronger evidence than any single source.
Most VoC tools are feedback databases with a search bar. vox is a research agent that applies real PM methodology:
- Thematic analysis (Braun & Clarke, 2006) — not keyword counting
- Jobs to Be Done (Ulwick, Christensen, Klement) — not feature request lists
- Opportunity mapping (Teresa Torres) — not priority matrices
- Evidence-based personas (Cooper, Goodwin) — not workshop assumptions
- Pain point severity scoring — not High/Medium/Low dropdowns
The methodology is in the /frameworks/ directory. It's comprehensive, opinionated, and based on the actual literature — not watered-down summaries.
vox is a Claude Code project — a collection of agent prompts, slash commands, framework docs, and output templates. There's no server, no database, no runtime. The intelligence lives in the prompts.
vox/
├── .claude/
│ ├── agents/researcher.md # Core research agent
│ └── commands/ # Slash commands (/voc, /persona, etc.)
├── frameworks/ # PM methodology reference docs
├── templates/ # Structured output templates
├── adapters/ # Data source setup guides
└── examples/ # Sample data and outputs
When you run a command, Claude Code:
- Discovers your connected data sources (MCP servers, local files)
- Scopes the relevant data for your question
- Applies the appropriate research methodology
- Produces structured output using the templates
- Cites evidence, quantifies findings, flags limitations
Any MCP server that exposes conversation data works. See adapters/custom.md.
- Write the methodology doc in
frameworks/ - Create a slash command in
.claude/commands/ - Create an output template in
templates/
Templates are markdown files with placeholder structure. Edit them to match your team's format.
See examples/sample-output/ for real example outputs generated from the sample data:
- Persona card — "Rachel: The Overwhelmed Insights Leader"
- Theme analysis — 5 themes from 12 customer conversations
- Opportunity map — Reducing enterprise churn
Customer research isn't optional and it isn't a luxury. Every product decision is a bet — research reduces the odds of betting wrong.
But most teams don't do rigorous research. Not because they don't want to — because it takes too long. Reading 200 transcripts takes a week. Building a proper opportunity map takes days. Scoring pain points across segments takes a full sprint.
vox compresses the time without compressing the rigor. The methodology is the same. The evidence standards are the same. The output quality is the same. The time is not.
The goal: any PM, at any company, can do world-class customer research in minutes instead of days. No enterprise contract. No training program. No research team required.
Contributions welcome. The most impactful areas:
- New frameworks — Kano model, empathy mapping, customer journey mapping
- New adapters — Salesforce, HubSpot, Zendesk, Intercom MCP setup guides
- Better prompts — The agent and command prompts are the core product. Make them sharper.
- Template improvements — Better output formats, more visualization
- Sample data — More realistic datasets for testing
MIT