codex exec --json for JSONL event streaming, which is significantly more automation-friendly. See Codex CLI Reference below./pack install business-discovery → /icp → then technical spike on Codex CLI. Market validation before engineering investment.A persistent Hetzner-powered remote development machine that runs multiple Codex CLI sessions in parallel, controlled from browser or phone. Five integrated surfaces — Agent Board, Terminal, VS Code, Browser Preview, and File Explorer — provide a complete AI coding workstation with zero public inbound access.
This is the personal dogfood environment. Agent Board + Terminal are the priority surfaces. VS Code, Preview, and File Explorer are deferred to existing tools.
Running AI coding agents locally is constrained by:
The sole beneficiary is you — a solo developer who wants to run 5-12 parallel AI coding sessions across multiple repos, monitor them from Mac and iPhone, and maintain a locked-down security posture.
Human-in-the-loop multi-agent orchestration from any device.
Agent Board + Terminal first. These are the two most critical surfaces. VS Code, Browser Preview, and File Explorer are deferred — existing tools (code-server, direct Tailscale tunnel) cover those needs until the core orchestration UX is proven.
| Constraint | Detail | Source |
|---|---|---|
| Provider | Hetzner Cloud US | Decided |
| Default machine | CCX33 (8 vCPU, 32 GB RAM) | Decided |
| Burst machine | CCX43 (16 vCPU, 64 GB RAM) | Decided |
| Scaling rule | Do not expand root disk on scale-up | Decided |
| OS | Ubuntu LTS | Decided |
| Network posture | Zero public inbound ports | Decided |
| Access layer | Both: Tailscale for direct terminal/SSH, Cloudflare Tunnel for web dashboard/PWA | Resolved Q2 |
| Agent CLIs | Codex CLI first (better programmatic API via exec --json); Claude Code deferred pending ToS review | Resolved Q3 |
| Session management | tmux-based | Decided |
| Browser IDE | code-server or OpenVSCode Server (deferred surface) | Decided |
| Code structure | Monorepo shared with commercial SaaS | Decided |
| # | Unknown | Risk | Notes |
|---|---|---|---|
| 1 | Codex CLI programmatic control | Critical | Technical spike before any UI. Codex has exec --json (JSONL events) and -o FILE. Spike should validate whether these provide sufficient control for the Agent Board. |
| 2 | Permission prompt interception | Critical | Codex has --dangerously-bypass-approvals-and-sandbox and -s MODE sandbox policies. May provide a path for web-based approvals. |
| 3 | Session state observation | High | Codex --json JSONL output may stream status events. Needs spike validation. |
| 3.5 | Claude Code ToS compliance | High | Potential terms-of-service concerns with programmatic control. Focus on Codex first; revisit after ToS review. |
| 4 | Output streaming | Medium | How to stream agent output to web UI in real time. |
| 5 | Resource ceiling | Low | Whether CCX33 handles 10+ parallel sessions. Testable empirically. |
Key Codex CLI capabilities relevant to the Agent Board:
codex exec "prompt" — Run a prompt non-interactively (alias: codex e)--json — Print events to stdout as JSONL (state observation, output streaming)-o, --output-last-message FILE — Write agent's final message to a file--output-schema FILE — Constrain final response shape with JSON Schema-s, --sandbox MODE — Sandbox policy: read-only, workspace-write, danger-full-access--dangerously-bypass-approvals-and-sandbox — Skip prompts (for externally sandboxed envs)--ephemeral — Don't persist session files--color never — Disable color output (for piping)-m, --model MODEL — Select model-i, --image FILE — Attach image(s)-p PROFILE — Select config profile (NOT equivalent to Claude Code's -p)
Note: Codex's -p selects a configuration profile from config.toml, not print/prompt mode. The equivalent of Claude Code's -p is codex exec. Stdin is also supported — if no prompt argument is given, instructions are read from stdin.
Resolved sequencing: Business discovery first, then technical spike.
/pack install business-discovery — install research skills for commercial concept./icp — validate beneficiary hypotheses (starting with power users / solo founders).codex exec --json — prove feasibility of programmatic control before any UI work./feature-interview on Agent Board + Terminal once CLI strategy is confirmed.A browser/mobile SaaS control plane for AI coding agents on gblock-party's managed cloud infrastructure. Users connect their preferred front-end tool (T3 Code, Multica, Claude Code CLI) or use the native Agent Board.
All infrastructure is managed by gblock-party on Hetzner.
Builds on the personal workstation codebase via a shared monorepo.
Developers increasingly use AI coding agents but face these problems at scale:
To be validated by ICP analysis. Other segments for later investigation:
Human-in-the-loop multi-agent orchestration from browser and mobile.
| vs. | Gblock Party advantage |
|---|---|
| Local CLI usage | Remote access, mobile, multi-agent visibility, session persistence |
| Browser IDEs (Gitpod, Codespaces) | Agent orchestration focus, not just "code in browser" |
| T3 Code | Better mobile UX, zero-inbound security, session management, diff review, approval workflow |
| Self-managed VPS + tmux | Turnkey control plane, no sysadmin, Agent Board UX |
| Constraint | Detail | Source |
|---|---|---|
| Build order | Personal workstation built and dogfooded first | Decided |
| Code structure | Monorepo shared with personal version | Decided |
| Commercial model | Hybrid SaaS + optional managed IaaS | Decided |
| Host OS | Ubuntu LTS (initial target) | Decided |
| Network posture | Host-initiated outbound only | Decided |
| Agent CLIs | Codex CLI first; Claude Code deferred pending ToS review; OpenCode later | Resolved Q3 |
| Managed infra provider | Hetzner (for optional IaaS tier) | Decided |
| No GitHub Actions | Project does not use GitHub Actions | Decided |
| # | Unknown | Risk | Category | Notes |
|---|---|---|---|---|
| 1 | Codex CLI programmatic control | Critical | Feasibility | Spike on codex exec --json. Must be resolved during personal dogfood. |
| 2 | Permission prompt interception | Critical | Feasibility | Can prompts be routed through web UI? |
| 3 | Session state observation | High | Feasibility | Real-time status from Codex JSONL output. |
| 4 | Daemon-to-SaaS tunnel | High | Architecture | Technology choice undecided. |
| 5 | Willingness to pay | High | Market | Pricing genuinely unknown. Defer to monetization research — no anchoring on preliminary ranges. |
| 6 | Data custody trust | Medium | Market | Trust barrier for tunneling through SaaS. |
| 7 | Daemon reliability | Medium | Engineering | Auto-updates, crash recovery on gblock-party's managed infrastructure. |
| 8 | Multi-agent CLI compatibility | Medium | Feasibility | Whether future agents have similar interfaces. |
Ready for ICP once /pack install business-discovery is run.
Resolved sequencing: Business discovery first.
/pack install business-discovery — install ICP, competitive analysis, value prop, positioning, lean canvas./icp — starting with power users / solo founders hypothesis.