Project Info
Inspiration
The Codex pet keeps me company while I work. I wanted to evolve that idea into a more realistic human presence: not another chat window, but someone who feels at home on the Mac and gives a sense of presence.
What it does
Baegent is a local-first, Unreal-native desktop companion for Apple Silicon Macs. Five Codex lifecycle hooks turn session, thinking, tool, and completion events into visible presence. A separate Model Context Protocol path lets Codex discover Baegent's authenticated native abilities, so the companion is not only an animated status indicator. The integration is content-minimizing by design. It forwards the lifecycle metadata needed for presence, but it does not forward or spool prompt text, source code, file contents, transcripts, tool inputs, tool results, assistant text, or credentials.
How we built it
Baegent builds on an earlier Poise foundation: an Unreal and MetaHuman avatar runtime, transparent macOS presentation, voice and facial-animation pipelines, and proof tools. That foundation predates Build Week. During Build Week, I used Codex and GPT-5.6 as the primary implementation partner to make Codex a first-class participant in the product. We added a native C++ agent-ability layer and authenticated loopback facade, then built the tvaily-codex plugin with five lifecycle hooks and a bundled MCP server. Swift and Python components handle privacy filtering, stable event identity, ordered delivery, admission receipts, a bounded offline spool, and targeted replay after Baegent returns. Codex worked in small red/green slices across C++, Objective-C++, Swift, Python, and Unreal. I remained the product owner and reviewer. Key human decisions included discarding content at the adapter boundary, keeping hook trust as an explicit user choice, using a native product path, and making the diagnostic tools fail when one proof level is mistaken for another.
Challenges we ran into
Conversational latency and facial animation add another piece to the real-time puzzle. We also had to make agent activity visible without turning the companion into a content-surveillance system. The hardest engineering challenge was reliable delivery across several independent boundaries. A hook process exiting successfully does not prove that the native bridge accepted an event. A green build does not prove that the packaged app ran. A playback marker does not prove that a judge heard audio. We built a deterministic doctor and evidence workflow that keeps source, hook trust, bridge admission, native MCP, packaged runtime, visible pixels, and audio as separate claims.
Accomplishments we're proud of
I am most proud that Tvaily is two-way and privacy-conscious. Codex can inspect the native ability surface and receive an authenticated result from the Unreal runtime, while the lifecycle adapter discards prompts and work product before transport or offline storage. The Build Week integration includes reversible install, doctor, pending, replay, evidence, and uninstall commands. Its source-facing proof can run without rebuilding Unreal. The core integration was developed with GPT-5.6 and is tied to a recorded Codex session and dated commits. Along the wider path to Tvaily, we also created a custom MetalFX upscaler and frame interpolator for Unreal on macOS, a native C++ agent harness modeled after Tau and Pi, and a fully local inference option. Also, the transparent background of the Baegent was a fundamental requirement that Codex nailed.
What we learned
Presence is a delicate balance to program.
What's next
Next, we will tighten the packaged flow, reduce speech-to-face latency, and expand the native ability catalog without weakening its capability firewall. The larger goal is a trustworthy desktop collaborator that can act, recover, and communicate clearly while keeping the user's work local and private.
This project did not link a GitHub repository.
Analysis
No indexed repository for this project, so there are no commit stats to show.
Technology
- C++Unchecked
- PythonUnchecked
- SwiftUnchecked
No repository was indexed for this project, so these Devpost claims have not been checked against code.
AI coding agents
No repository was indexed, so agent usage could not be checked.
Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.
Codebase size
No repository was indexed, so there is no codebase to measure.
This project did not link a GitHub repository, so there is nothing to diagram.
This project did not link a GitHub repository, so its feature claims have not been checked against code.
Export this project's context (description, README, evidence, key source files) to chat with an AI agent elsewhere.