M.I.R.A. v2

People in marketing and e-commerce jump every day between Shopify, Klaviyo, Meta, documents, briefs and code, and put the numbers together by hand.

I needed one place to ask a question and get an answer that's already reasoned through, with the data visualised.

A router reads the request and passes it to the best agent: Marketing, Research, Development, Reports, Assistant, and Soolù in the personal version.

When it talks numbers it draws charts: lines for trends, bars for comparisons, never with invented figures. It reads PDF, Word, TXT and CSV, speaks and listens, remembers recent exchanges and shows how many tokens it used.

Orchestration is an n8n workflow: Claude Haiku is the router, Claude Sonnet answers with the chosen agent's prompt. Each agent is a Markdown file versioned on Git, and a Python script regenerates the workflow from those files.

Accounts and projects on Supabase with Row Level Security. The app is plain HTML, CSS and JavaScript, installable as a PWA, with hand-made SVG charts, a Canvas voice wave and in-browser document reading with pdf.js and mammoth.

Security first: Telegram answers only me, accounts are invite-only, secrets live only in the backend, and documents and emails are treated as data, never as instructions. The demo has server-side usage limits. Over 90 automated tests, Playwright interface checks and a check that blocks keys and tokens in the code.

Keeping prompts in separate files, one per agent, makes the system easy to change and test: I edit a file and regenerate the workflow.