Praxis AI Agent Platform — Data skill for Claude Code

Praxis

An autonomous enterprise data and operations intelligence platform.

You give it a question in plain language. It plans the work, runs that plan against your real systems under your real permissions, pauses for your approval before anything that changes state, and hands back an answer with the evidence behind it. If a step fails, it reads the error, re-plans, and tries a genuinely different approach.

What it does

• Answers questions over your data. Point it at a warehouse, an API, a metrics server or a spreadsheet you just uploaded. It works out which to use and how.
• Computes exact numbers. Uploaded spreadsheets become typed tables queried with real SQL, so "what is the total" returns arithmetic, not something that reads like an answer.
• Builds new capabilities when none fits. It writes the tool it needs, validates it in a locked-down container, and refuses to run it until a human approves that exact code.
• Investigates open questions. For asks that cannot be planned ahead ("which region is underperforming and why"), it runs an observe/decide/act loop and shows every step it took.
• Never changes anything behind your back. Mutating actions stop and wait for a person.

How this helps you

You want to	Praxis gives you
Ask across systems without writing glue	One intent; it picks the connectors and composes the steps
Trust the number	Exact SQL over typed data, plus the query and rows it used
Let agents act, safely	Approval gates bound to the exact arguments and code approved
Know what happened	Per-step evidence, an audit row per decision, full replay
Not babysit failures	It diagnoses the error and re-plans instead of stopping
Avoid lock-in	Swap Anthropic/OpenAI models with one env var; add a connector in one file

---

Running it

Requires Python 3.11, Docker, and a POSIX-ish shell.

1. Start the infrastructure

docker compose up -d          # P