Aurolabs · workshop

Vibecoding, agents & tools for pros.

A small workshop. Vibecoded, shipped, used daily.

/ 01

Flagships

Products with their own surface, their own users, their own pricing.

flagship · private preview

The flagship. An agentic decision platform — agent-capable from the first message, adversarial review baked in, briefs that keep watching.

What: A full agent inside every session, plus a four-stage decision pipeline: framing chat learns your decision; adversarial research runs two top models against each other for 5–30 minutes; the brief lands as a card you can challenge, extend, or set on watch over time.
Why: Most AI research dies at the brief — read once, filed, forgotten. Auroagent keeps the answer alive as an artifact you can interrogate, runs adversarial review at every stage, and ties it to ongoing signal so it stays current. Decision-grade, not search-grade.
How: Wraps the Quanto adversarial engine. Magic-link auth, multi-session memory, decision-card brief delivery with linked § citations, cron-driven watches with live web search and email alerts. Built for decisions that need to survive scrutiny.

side-threads · adversarial review

A native macOS chat for Claude Code with parallel side-threads and on-demand adversarial review from a second model.

What: A desktop UI that spawns the claude CLI and reads ~/.claude/projects locally. Side-threads let you branch a sub-conversation in the side panel without losing the main thread. With ChatGPT Bridge installed, any answer can be sent for adversarial review by a second model — landing back in the same session.
Why: Claude.ai is already clean. The leverage is parallel context — explore a tangent in a side-thread without losing the main conversation, and pull a second pair of eyes when the answer matters.
How: MIT, Tauri + React + Rust. ChatGPT Bridge ships from the same shelf. Optional Aurochat Pro at aurolabs.ai/chat adds zero-knowledge cloud sync across devices.

topic watcher · email digest

AI that keeps an eye on the things you care about, and quietly pings your inbox only when something actually changed.

What: One brain-dump in. Pssst splits it into up to three focused topics, picks your rhythm — monthly, weekly, daily — and watches them in the background. When something genuinely new shows up, a short, citeable digest lands in your inbox.
Why: Newsletters drown you, manual search is work, and most research agents over-talk. The leverage is a quiet, opinionated agent that only speaks when there is something worth saying — and ships you the link before anyone else has reposted it.
How: Node + Express, magic-link auth, cron-driven multi-search orchestrator with source verification. Monthly forever-free, weekly €9/mo, daily €29/mo. Public /v1 API at €0.50 per scan + five free for agents that need their own pair of eyes.

/ 02

Sharp utilities, open engines, and a shipyard full of agents.

Claude ↔ ChatGPT · MCP

Pipes Claude (or any MCP host) to a live ChatGPT desktop session and reads back the reply.

What: Ships as both a Claude Code plugin and a stdio MCP server for Claude Desktop. cgpt-send / cgpt-read-app / cgpt-read-share / cgpt-wait-idle / cgpt-attach — all exposed as MCP tools.
Why: When two top models can argue, you get answers neither one would give alone. And once it speaks MCP, any agent host can use it.
How: Swift shebang scripts drive the ChatGPT desktop app via macOS automation; a small Python stdio MCP server wraps them as tools, plus a launch-bridge skill for orchestration.

Deep Adversarial Research · request access

Two top-tier models researching against each other until the brief survives the loop.

What: A Director (GPT-5.4) challenges a Researcher (Opus 4.7) for N iterations. Every cited URL gets re-fetched and verified before the brief ships.
Why: Single-model research hallucinates. Adversarial loops kill claims that cannot survive a second pair of eyes.
How: Opus does the research turns, GPT directs and challenges, source-fidelity pass re-fetches every URL, brief gates on convergence.

every Factory ship

Where Aurolabs Factory ships its agents — new verdicts, new builds, new wedges.

Vibecoded in Stockholm by a guy named Roberto, on paper-feel pixels and two arguing models.

— Aurolabs · 2026 ·