This Is What It Looks Like
You talk to an AI agent. It uses Praxis behind the scenes — searching structured knowledge, incorporating research, running analyses, and comparing results to the literature. Here's a real conversation:
We're seeing higher churn in our mid-tier customers. I want to understand what the research says about predicting and preventing customer churn. What do we know?
Let me check what analytical frameworks we have for this.
Good news — we have the customer-segmentation-rfm PAX installed, which covers RFM analysis for segmentation. But I also see a gap: we don't have a dedicated churn prediction PAX. Let me check the marketplace for one.
The credit-risk-ml PAX shares relevant constructs — both domains use logistic regression and random forest on binary outcomes with behavioral features. Its prediction_accuracy and false_positive_rate constructs directly apply to churn modeling. Want me to explore what we can pull together?
Yes — and while you're at it, I just read a paper about how customer satisfaction links to retention. Can you incorporate that into our knowledge base?
Sure, share the paper and I'll extract the key findings into our KB.
Done. I extracted 3 constructs and 4 findings from the paper:
customer_satisfaction— overall satisfaction score (positive → retention)switching_barriers— contractual/procedural lock-in (positive → retention, but weaker)affective_commitment— emotional bond with brand (strongest individual predictor)
Key finding: affective commitment mediates 60% of the satisfaction-retention link. Customers don't stay because they're satisfied — they stay because they feel connected. This changes the intervention target from "make them happy" to "make them feel belonging."
That's a great insight. Now can you build a churn prediction model using our customer data and test whether satisfaction or commitment is the better predictor?
Logistic regression on 12,400 customer records:
affective_commitment: β = -0.71 (p < 0.001) — strongest predictor of retentioncustomer_satisfaction: β = -0.12 (p = 0.08) — not significant after controlling for commitmentrecency: β = +0.54 (p < 0.001) — longer since last purchase = higher churn riskKB Comparison: Your result confirms Gustafsson et al.'s finding — commitment dominates satisfaction as a retention driver. Your recency effect is consistent with the RFM literature. Recommendation: focus retention interventions on community and belonging programs, not just satisfaction surveys.
This is great. Package everything we just did into a PAX so the rest of the analytics team can use it.
Done. Created customer-churn-prediction PAX with 5 constructs, 4 findings, and a quick-start playbook that reproduces your analysis. Anyone on the team who installs this PAX gets your constructs, the literature findings, and the playbook — they can run the same analysis on new data without starting from scratch.
The full loop: ask a question, search the knowledge base, incorporate new research, run an analysis with automatic literature comparison, and package everything into a reusable PAX. All in natural conversation.
Set Up Your Own Instance
Ready to try it? Here's how to get Praxis running and connect it to your agent.
01 Install Praxis
git clone https://github.com/JELambert/praxis.git
cd praxis && uv sync
uv run python -m praxis.mcp_server02 Connect Your Agent
Add Praxis to your Claude Code MCP config:
{
"mcpServers": {
"praxis": {
"command": "uv",
"args": ["run", "--directory", "/path/to/praxis",
"python", "-m", "praxis.mcp_server"]
}
}
}Works with any MCP-compatible agent — Claude Code, Codex CLI, or custom agents via the Agent SDK.
03 Install PAX from the Marketplace
curl -O https://pax-market.com/packs/happiness-economics.pax.tar.gz
praxis_import_pax("happiness-economics.pax.tar.gz", install=True)Or just ask your agent: "Install the happiness-economics PAX from the marketplace."
04 Start Talking
That's it. Ask your agent anything about the domains you've installed. It has 73 Praxis tools available — knowledge search, cross-domain pathways, 18 analytical engines, paper incorporation, and PAX creation. You saw how it works above.