Critias

Critias is an open-source pricing agent: a local browser application for AI-assisted pricing and gross margin analysis. Ask questions about CSV, Excel, text and Markdown files and get answers whose figures are checked against the code that produced them, with native tables and Plotly charts read straight from the computed results.

Built on Pi (earendil-works/pi), the open-source agent harness. Pi supplies the model layer, the agent loop, and tool calling; this project adds the pricing skills, the data checks, and the figure verification on top.

The application combines an assistant-ui chat interface (React), a Fastify server, the Pi agent SDK, DuckDB, seven bundled pricing skills (Pricing Plugin 0.4.13) and a visualization skill. Calculations run locally. Model requests go to the provider you choose: OpenAI, OpenRouter, Google Gemini or Anthropic.

Status: experimental research preview (0.5.1). Native macOS setup, component, Python and browser tests, and live runs of the single-analyst workflow on synthetic and public datasets have been checked. Real-business-data validation and native scale testing are still pending. See how the harness works and the verification records.

Features

• One continuing analyst session per conversation. It works in SQL over typed views of your files and in recorded Python scripts, and follow-ups keep the population, period and basis agreed earlier.
• Seven pricing skills: intake, margin analysis, opportunity scan, recommendations, pricing design, rollout planning and triage, with knowledge and method cards.
• Checked replies: every substantial number must appear in the output of a query or script that ran. Tables and charts are built from frozen result files, never from model text.
• Data findings: before the first question on a set of files, seven code-based detectors flag suspect rows (formats, spelling variants, missing keys, duplicates, broken products, date order, values far from their peers). The analyst asks how to treat them; nothing is cleaned automatically.
• Confirmed business context in .pricing/context.md, written only for facts you confirm, and conversation history in a sidebar.
• Local chart and table interactions (hover, zoom, search, sort) with no model requests, and live step progress while a question runs.
• Per-question limits on time, model calls and tool calls, cancellation, and connection recovery.
• A headless interface for scripted questions and a graded evaluation loop.

Skills and dependencies are pinned. The provider and model are set in pricing-harness/.env; the app never switches models or updates its bundled skills by itself.

Requirements

• macOS or Windows for the supplied setup and launch scripts. Windows also needs Git for Windows: the analyst runs its shell commands in Git Bash.
• Node.js 22.19 or later and Python 3.12.
• An API key and a model ID for OpenAI (or a stored ChatGPT login), OpenRouter, Google AI Studio (Gemini) or Anthropic. Any model ID works, including ones newer than Pi's catalogue.
• Internet access for installation and model requests. Chromium is only needed for the development browser tests. Docker isn't needed.

Install with npm

You need Node.js 22.19 or later and Python 3.12 (Windows also needs Git for Windows).

npm install -g critias

Create ~/.critias/.env (on Windows, %USERPROFILE%\.critias\.env) with a key and a model for one provider. For example:

ANTHROPIC_API_KEY=your-key
ANTHROPIC_MODEL=claude-sonnet-5-5

The other providers are OPENAI_*, OPENROUTER_* and GEMINI_*; critias models anthropic lists model IDs. Then start the app with a folder for your business files:

critias ~/CritiasWorkspace

The first start creates a Python environment in ~/.critias/venv, which takes a minute and needs a network. Open the link it prints. npx critias works without a global install. To update, run npm install -g critias again.

Quick start from a clone

Run these commands from the repository root:

cd pricing-harness
bash scripts/setup-macos.sh
cp .env.example .env

Set a key and a model for the provider you use in .env: OPENAI_API_KEY and OPENAI_MODEL, OPENROUTER_API_KEY and OPENROUTER_MODEL, GEMINI_API_KEY and GEMINI_MODEL, or ANTHROPIC_API_KEY and ANTHROPIC_MODEL. The app uses the first provider that has both, in that order. With a stored ChatGPT login (Pi's ~/.pi/agent/auth.json), set only OPENAI_MODEL. Keep the file out of Git. Then launch with a separate business workspace. An optional second argument overrides the model for one launch:

bash scripts/start-macos.sh "$HOME/PricingHarnessWorkspace"

Open the local application. The launcher creates the workspace if needed. Restart after changing harness.config.json or .env; the thinking level is chosen in the interface, beside Send; stop with Ctrl+C. Model requests use your credentials and can incur provider charges.

Add a file and ask, for example: “Which products contributed most to the change in gross margin between these years?” Confirm column meanings, currency, calendar, cost treatment and comparison scope when asked. See local setup for model selection, direct Node and OpenAI launches, headless use and every setting.

This is the beginning of the README. Read the rest at the original. Read the full README ↗