Auditable AI workflows

What we build

Many teams produce the same reports every month from spreadsheets and exports. A language model can do much of that work, but in a report that people sign off, each number has to be checkable. Our MCP (Model Context Protocol) tools are built for this kind of work. They run inside Claude and other assistants. In them:

• a language model plans the steps;
• tools do the computing, so the model does no arithmetic;
• every figure traces to the source row it came from;
• the result is checked by a separate step before it leaves;
• when the system is unsure, it stops or asks.

What Praxis does

Praxis is our engine for these workflows. It runs as a local MCP server on your machine.

• What goes in. Your files (CSV, XLSX, JSON or plain text) and your statement of which columns mean what.
• What comes out. A report in which each figure links to the source rows it came from. The arithmetic is done by code in exact decimal arithmetic, not by the model.
• When an input is unclear. Praxis stops or asks a question.
• Next month. A report set up once is saved and can be run again on new files without calling the model.

Version 0.1.0 is dated 26 October 2026. It comes as a single-file build for Linux.

Data and security

• Praxis runs on your machine. We do not hold your data.
• When it is used inside an assistant such as Claude, the assistant sees the tool results under your own agreement with its provider. Where that is not acceptable, we design the workflow around a model you host.
• We sign an NDA before you share any file. Data-processing terms are agreed per project.
• No security certification yet. We answer your security questionnaire.

What we measured

On 116 jobs built from real public data, no result that Praxis marked as certain was wrong.

The comparison is with the model answering alone, without code. On your own task we measure the approaches you are choosing between (next section).

Measuring on your task

We built a measurement harness that compares any reporting approach (a model alone, a model with tools, or a template) against an independent answer key, and records what each one costs.

We run it on your reporting task before we build. If the numbers say a model writing its own code, or a plain template, does your job better, that is what we recommend.

Limits today

• Preview. Praxis 0.1.0 is a preview release at alpha stage. It is not a public online service, and its interfaces may change.
• Platforms. The public build is for Linux. For a client we set it up on Windows or whatever environment they use.
• Input formats. Currency and date formats are being extended; we check yours in the first step.
• Questions. On real jobs Praxis asked more questions than needed. This is partly fixed.
• Licensing. There is no licensing or payment set-up for Praxis as a separate product yet.

How a project starts

You send one recurring report, the files it is built from, and a past edition to check against. We reply within one working day and say whether the report suits this kind of automation. The first job covers that single report and usually takes one to three weeks. Scope and price are written down and agreed before it starts, and nothing is owed until you accept the result. Further reports are quoted once the first is accepted.

What you receive

• The report itself, with each figure linked to the source rows it came from.
• The measurements of the candidate approaches against your past edition, with what each one costs.
• Our recommendation from those numbers: Praxis, a model writing its own code, or a plain template.
• The report set-up, saved so it can be run again on new files.

If the report's numbers do not come from files you can export yet, tell us where they come from and we will say what to export or start recording.