A second set of eyes for Codex claims
Agents can make confident claims that later drive emails, webhooks, deployments, or other actions. The person using the agent often has to choose between trusting the answer and manually checking every detail.
gBox works alongside ordinary Codex tasks. After a response completes, it extracts material claims, selects an eligible read-only MCP or web source, retrieves evidence, and reports each claim as Verified, Contradicted, or Unverifiable. A native notification and optional macOS camera-notch surface bring exceptions forward without replacing or interrupting Codex.
Review opens a dossier showing exactly what gBox understood, which source it selected, the raw evidence, comparison method, and any failures. Contradictions can be copied back as an evidence-backed correction. When the bundled protected webhook is requested, gBox requires explicit human approval and records the claim, evidence, decision, and exactly-once result in a local hash-chained receipt.
How it was built
gBox is a macOS-first Tauri application with a Rust core, React and TypeScript interface, SQLite ledger, real Codex App Server supervision, trusted Codex plugin hooks, and MCP integrations. Passive observation is non-blocking; the protected action fails closed. The verifier is generic: it plans over eligible MCP tools or web sources rather than being tied to the seeded company example.
Codex and GPT-5.6
Codex with GPT-5.6 was the primary build environment. It helped design the trust boundaries, implement the Rust and TypeScript integrations, create regression coverage, iterate on the interface from screenshots, and run the live macOS end-to-end demonstration. In the demo, Codex uses computer control to prepare gBox, opens an ordinary task, produces a deliberate false claim, and leaves that task visible while gBox reports the contradiction.
Try it
Apple Silicon judges can download the runnable macOS release without a development toolchain: https://github.com/astinz/gbox/releases/tag/v0.1.0
Unzip it, drag gbox.app to Applications, Control-click it and choose Open on first launch, then choose Test tools → Deterministic replay. Replay requires no Codex authentication, credentials, source build, or external network. It exercises all three verdicts, the real approval dialog, loopback delivery, persistence, and receipt verification.
V1 observes local Codex surfaces that trust the gBox plugin and protects only the fixed bundled test webhook. Company data is synthetic, and receipts are locally hash-chained rather than externally notarized.
Built With
- codex
- gpt-5.6
- mcp
- react
- rust
- sqlite
- tauri
- typescript