DON BUG GPT
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

INTI AI is a local-first desktop agent that gives one operator an auditable control surface for work performed on a Windows PC. A bilingual Spanish/English dashboard pairs with an agent on the operator's machine. From it the operator can: - Pair a computer with a one-time code, inspect installed applications, grant or revoke per-device permissions, and see device state. - Capture an allowlisted web viewport and, with a configured Meta key, have Muse Spark turn it into a typed scene of elements, UI state and uncertainties. - Convert a goal and constraints into a typed plan with ordered actions, targets, preconditions, expected results and risk levels; unsafe proposals are rejected and planning never authorizes execution. - Approve the exact plan by SHA-256 digest, with separate confirmation for high-risk actions; any edit invalidates approval. - Execute approved actions in the synthetic Chromium inventory sandbox or on the local PC: create and read back text files, request Notepad or an approved HTTP(S) page, and launch authorized applications without arguments. Progress arrives as live events and per-action receipts; browser sessions pause and resume, remaining work can be cancelled, and history exports as JSON. - Generate a document dossier in a fresh task folder (Word quotation, Excel budget with live formulas, native chart and reconciliation sheet, print-ready PDF, text summary and SHA-256 manifest), then reopen it to check fields, totals, structure and hashes. - Diagnose failed sandbox executions with a bounded repair plan that requires fresh approval and never replays actions automatically. - Verify sandbox outcomes independently: re-read persisted state, evaluate explicit criteria, collect hash-checked evidence, and issue VERIFIED_SUCCESS, VERIFIED_FAILURE or INCONCLUSIVE with JSON/HTML reports. - Review measured latency, throughput and verification results, with unavailable metrics left blank instead of estimated. It is a single-operator local release. The agent writes only inside the approved workspace and task folder, never overwrites, signs or sends files, and exposes no shell, deletion or arbitrary-executable command. Desktop screenshot capture, independent desktop-vision verification and automatic model calls are not implemented or disabled by default.