Tora AI

A local-first AI agent you text like a person through Telegram. From your phone it searches the web with sourced answers, reads feeds with proof screenshots, manages files, and fills out multi-step job applications from a saved profile and resume — with the reasoning running on a local model, no cloud API anywhere in the loop. Anything irreversible triggers an Approve/Deny card on your phone showing exactly what's about to happen.

• ▸Human-in-the-loop safety architecture: deterministic code-based rules — never model judgment — gate 100% of irreversible actions behind phone Approve/Deny cards showing the exact values involved; every action is written to an append-only audit log.
• ▸Approvals resolve asyncio Futures, suspending the agent mid-task until a human responds.
• ▸End-to-end ATS job applications with a browser-use agent that reads live page state instead of guessing CSS selectors; a custom click-final-submit tool means submission requires fresh human approval every time; a JSON job store guarantees zero duplicate applications.
• ▸Hardened tool-calling loop for small local models: lenient JSON extraction, repeat-call guards, and error-as-tool-result recovery so no tool failure crashes the agent.
• ▸Domain-scoped credential placeholders — the model sees 0 real secrets. Playwright pinned to a dedicated thread for Windows async stability.
• ▸Policy: never solve or evade bot checks — a paused agent is correct behavior, not a failure.