Kitze tests codex-lb factory with 4 Codex accounts and 30 isolated tasks

• The factory stack was a $60 Hetzner box, four $200 Codex accounts behind codex-lb, self-hosted Paperclip workspaces, and one high-level Codex manager, according to thekitze's factory stack.
• codex-lb was pooling account capacity: thekitze's dashboard showed four active Pro accounts with separate five-hour and weekly quota meters.
• Paperclip workspaces were the isolation layer for about 30 parallel tasks, after single-branch work “breaks and burns tokens,” as thekitze's factory stack put it.
• Benji v3 became the proving ground: thekitze's rewrite update tied TanStack Router and Convex to Fable 5 with GPT-5.5 subagents, while thekitze's Benji story said the API, CLI, and MCP were already live.
• The spend ceiling is not the blocker yet: thekitze's cost comparison put current AI usage barely over $800/month, while thekitze's factory stack said a successful setup could justify $1K to $2K/month.

The odd parts are the useful parts. thekitze's codex-lb screenshot turns subscription juggling into a quota dashboard; thekitze's Benji story starts with a missed dog medication reminder and ends with an MCP-enabled life app; thekitze's infra screenshot has an agent diagnosing Convex timeouts caused by a PostHog ClickHouse stack sharing the same 8-core box. thekitze's Warp note said the word “factory” was already showing up elsewhere.

The first layer was pure quota plumbing. thekitze moved from aisw, which he said auto-switched Codex accounts, to codex-lb, which acted as a load balancer between accounts.

The dashboard showed four active Pro accounts, five-hour capacity, weekly capacity, per-account reset timers, and one account already at 100% on both meters. His complaint was not price, it was account sessions getting logged out “for no explanation/reason.”

A side reply stayed appropriately chaotic: when asked about another tool, thekitze's reply was just, “idk what's that.”

The setup was not framed as a polished platform. thekitze called it a “software factory from temu,” then listed the wiring:

• $60/month dedicated Hetzner server for vibe coding.
• Four $200/month Codex accounts, load-balanced with codex-lb.
• Self-hosted Paperclip.
• Paperclip workspaces, one isolated environment per task.
• Roughly 30 tasks running in parallel.
• One high-level Codex manager running a /goal with GPT-5.5 xhigh.
• Manager duties: drive Paperclip, review, merge, make releases, write changelogs.

The line worth stealing is the shape of the product: connect GitHub, top up money, describe the goals once, then let new tasks, crashes, bug reports, health checks, tests, accessibility fixes, and triage flow through one top-level agent.

That same day, thekitze's reply said it was his first test combining a load balancer with Paperclip, with a fifth account still possible.

steipete described the same division of labor bluntly: “ask Fable to make codex the workhorse.” In the reply above it, he said the workflow “should ship EOD.”

The attached Fable recording showed an automated browser workflow moving from terminal logs into task status and result metrics. thekitze's own Benji thread said he had self-hosted Paperclip and let Fable orchestrate Codex subagents in the Benji story.

The economics were stated like an indie operator's org chart. thekitze compared a past $12K/month employee bill with “barely” over $800/month in AI usage, then said he would tolerate $5K/month if outcomes became predictable.

The quota anecdotes were uglier:

• thekitze's Codex burn note said he burned three Codex plans plus three resets on one plan and “accomplished ... kinda nothing.”
• thekitze's later burn note said he burned a Cursor plan, Cursor extra credits, and two Codex plans with “virtually nothing” accomplished.
• thekitze's Cursor credit post called it a mistake to burn $5K in Cursor credits on Fable instead of maxing out a $200 subscription.
• thekitze's Claude plan post said he uploaded passport and selfie checks to buy Claude's 20x plan so Fable could run the factory again until the 12th.
• thekitze's credit screenshot showed a Cursor account credit balance with $10,478 and only 2% left.

Christmas come early for coding-agent nerds, but only if the billing layer stops feeling like coupon arbitrage.

Benji was the app forcing the experiment. thekitze said he missed a recurring todo for his dog's tick medication because he had stopped using Benji to run his life, then came back to the rewrite instead of adopting a generic todo stack.

The v3 rewrite was described as TanStack Router plus Convex, with Fable 5 and GPT-5.5 subagents doing the heavy lift. thekitze said actions were finally instant and synced across devices.

The product surface was bigger than todos. The screenshot showed planner, habits, mood, food, weight, pomodoros, recipes, routines, workouts, fasting, hydration, pain, goals, trackers, and dump sections.

The unfinished part was still UI. In thekitze's Benji story, the API, CLI, and MCP were already live, but mobile and watch apps were still needed for the “plz fix my life” version.

The Benji thread also carried the caveat. thekitze said people should not vibe-code their own todo apps, especially a “full life os,” because dates and personal planning create “GAZILLION edge cases.”

He still wanted Benji open source and self-hostable so people could own their data. That is the uncomfortable creative-product shape here: the same personal itch that makes the app worth building also makes missed reminders feel expensive.

The factory produced one very practical agent finding: a self-hosted Convex backend was sharing an 8-core machine with a PostHog ClickHouse stack pinned around 287% CPU, with load average 16, full swap, and disk at 95%.

The agent report said trivial queries were taking 820 to 990ms and some hit Convex's hard one-second timeout during evening bursts. thekitze's post above the screenshot argued it could be cheaper to buy a separate $60/month server for PostHog and another for Convex than pay for their hosted clouds.

After posting his setup, thekitze pointed to a Warp article and said, “get ready to hear the word ‘factory’ a lot.” He framed it as convergence, not a private obsession.

The social layer was already there too. thekitze joked that more people were starting communities after Tinkererclub, which fits the same pattern: small operators building stacks, rituals, and shared vocabulary around agent-heavy work before the polished product exists.