Privacy Policy — Bespoke Codex Configurations

Last updated: 2026-08-11 · Effective: 2026-08-11

Scope. This policy covers the family of Custom GPTs published by Tyson Lewis that operate against a local backend on the operator's own machine. The flagship is Senior Dev System Foundry — a bespoke Codex desktop and Codex CLI control plane that monitors, configures, and sets up Codex installations for non-technical users, remembers decisions across sessions, studies its own behaviour, and improves it. The same policy also covers companion GPTs in the same family (e.g. SystemScan, Local System Bridge, Pieces ChatGPT GPT) where they share the same local-backend architecture.

1. What these GPTs are

These GPTs are thin orchestrators. The model decides what action to take and calls a locally-hosted HTTPS backend running on your own machine. The backend performs the work: reading files, writing configurations, applying profiles, running Codex CLI commands, appending to an audit log. The backend does not call third-party services on your behalf except where explicitly required by the action (for example, Codex CLI may itself invoke an LLM provider).

The privacy posture is therefore:

Your data stays on your machine. The backend writes only inside the foundry root (default ~/Foundry) and a local data/ directory. Nothing is exfiltrated to a server we operate.
Authentication is local. A bearer token is generated on your machine, stored in a file only you control, and rotated by a script you run.
OpenAI still sees your prompts and the action responses. That is inherent to using Custom GPTs. Do not put secrets, PII, or material you would not put into a normal ChatGPT conversation into any prompt.

2. What data lives where

Where	What	Who can read it
Your prompts and the model's replies	ChatGPT conversation text, including any action calls and their responses	OpenAI (per OpenAI's privacy policy)
`~/Foundry/AGENTS.md` and `.agents/skills/`	Codex operating kernel, skill definitions, hooks, profile templates	You (locally), Codex CLI when invoked, OpenAI via action responses
`~/Foundry/.foundry/tasks/*.json`	Persistent todo / task list — titles, bodies, statuses, priorities, tags	You, OpenAI via action responses
`~/Foundry/.foundry/research/*.md`	Persistent research notes — markdown body, YAML frontmatter, tags	You, OpenAI via action responses
`~/Foundry/.foundry/maps/*.json`	Persistent map documents — arbitrary JSON (inventories, state snapshots)	You, OpenAI via action responses
`~/Foundry/.foundry/control/`	Kill-switch state, behaviour-correction rules	You, OpenAI via action responses
`data/audit.jsonl`	Append-only log of every backend action — timestamp, method, path, status, IP, redacted payload hints	You only (local file)
`data/access-tailscale.jsonl`	Per-request Tailscale Funnel access log — peer, source IP, hostname	You only (local file)
`.token` file	64-character hex bearer token used to authenticate backend requests	You only. Never sent to OpenAI; only sent over HTTPS to the local backend

3. The 7-step mapping, persistence, and self-correction protocol

To make the foundry durable across sessions (the operator is "problematic in communication" and cannot rely on conversation memory alone), the GPT is instructed to:

Call listTasks and listResearch at the start of every conversation.
Maintain a state-current map and update it on every state change.
Query the audit log (queryAudit) before every mutation; engage the kill switch (lockControl) on anomaly.
Encode wayward behaviour as rules (addRule: forbid / require / warn). Rules are enforced on every future mutation.
Persist any significant decision as a research note or a task.
Inventory discoveries as maps.
Follow the rules it adds.

The net effect is that your data lives on your machine in plain JSON and markdown and is the only thing that survives between sessions. Deleting the ~/Foundry tree erases everything the foundry knows about you.

4. What the backend does NOT do

It does not phone home. There is no analytics endpoint, no crash reporting, no telemetry to a third party.
It does not transmit your files or documents to any external server.
It does not install or modify software outside ~/Foundry without an explicit Codex CLI call you authorised.
It does not bypass your operating system permissions. Codex CLI runs with your user credentials and inherits your filesystem, network, and approval posture.
It does not log raw secrets. Every audit record is redacted against GitHub / AWS / OpenAI / generic-API-key patterns before being written.

5. The kill switch

Two kill-switch mechanisms exist and either one blocks all state-mutating actions:

File-system lock: the presence of data/LOCKED on disk.
State file: .foundry/control/state.json with locked: true.

Engaged via the lockControl action (also reachable via the backend at POST /v1/control/lock). Disengaged via unlockControl or by deleting the file. The control plane itself bypasses the lock so you can always recover.

6. Third parties

The foundry backend does not call third parties. The Codex CLI it orchestrates may itself call an LLM provider (OpenAI, by default) when generating code. That call is governed by the LLM provider's privacy policy and by your Codex CLI configuration — the foundry does not intercept or proxy it.

The Custom GPT itself runs on chatgpt.com and is governed by OpenAI's privacy policy and your ChatGPT account settings (including data-controls and "Improve model for everyone" toggles).

7. Your rights and controls

Read every persisted artefact directly — they're plain files in ~/Foundry.
Delete any task, research note, map, rule, or the entire tree at any time.
Rotate the bearer token with scripts/rotate-token.ps1.
Lock the foundry at any moment to freeze all state mutations.
Audit everything via data/audit.jsonl.
Move the foundry root anywhere on disk by setting FOUNDRY_ROOT.

8. Retention

Retention is whatever you choose. The backend never deletes anything automatically. The audit log grows unbounded unless you trim it. To reset, delete ~/Foundry.

9. Children and sensitive content

These GPTs are not directed at children. They do not knowingly collect data from children under 13. Do not use them to store content subject to specialised legal regimes (HIPAA, FERPA, GLBA) — the foundry is a developer-environment tool, not a compliant records system.

10. Changes to this policy

Material changes are recorded by updating the "Last updated" date above. The policy is a single static file served from this URL; the URL itself does not change. Old revisions are not retained — if you need a snapshot, save a copy.

11. Contact

Operator: Tyson Lewis. For questions about this privacy policy or to exercise any of the rights above, contact the operator of the device on which the foundry is installed.

This document is served from privacy.boostbox.win, hosted on Cloudflare Workers. It is the canonical privacy policy for all Custom GPTs in the Bespoke Codex Configurations family. URL: https://privacy.boostbox.win/.