I have spent nine years in B2B SaaS growth: go-to-market, SEO, content, paid ads, lifecycle marketing and product analytics. Over the past year I added AI agents to those same channels, so a channel can scale on compute.
Today I run dozens of Claude Code sessions in herdr on a Windows PC and Codex lanes on a Mac mini and VPS servers, across several Claude and Codex subscriptions. I built a few tools that make a fleet that size easier to run, and a checker for the copy it writes.
| Repository | What it does | Stack |
|---|---|---|
| anti-slop-content | Rules, a checker and an eight-step review line for marketing copy written by AI agents. Nine rules cover what agents get wrong: product claims, prices, compliance promises, setup times, competitors and unsourced numbers. Each rule is a data file with its source of truth, regex cues and test cases; an agent from a second model family tries to refute the facts, and a release ships only if its hash matches the last reviewed text. Everything about your company lives in one project/ folder. |
Python, standard library only |
| herdr-sidebar | herdr plugin for a Spaces sidebar of twenty-plus workspaces full of AI agents: titled categories with drag and drop, four kinds of colour stars with a hotkey to cycle through each kind, jump hotkeys to a project or a single tab, and git worktrees detached into a category of their own without stopping their agents. Agents on 24/7 duty check in with herdr-duty; one that stops waking up, hangs on a question or loses its pane turns red and can ping Telegram. |
Node.js, no dependencies |
| subtrack | Local dashboard for a fleet of Claude Code and Codex windows on several Claude, Codex and Grok subscriptions. Shows what is left of every five-hour and weekly limit and which sessions burned it, ranked by a cost-weighted token count; Codex sessions on other machines are read over ssh. Lists every Claude window in herdr, longest idle first: one click opens it or moves it to another subscription, and a mode per window decides whether its prompt cache is kept warm or the window gets compacted. | TypeScript, Node.js |
| teammate | CLI that lets one Claude Code session hire another as a worker in its own herdr tab: a written brief, a status log the worker appends to, and a close command that refuses to discard uncommitted work. Each worker can get its own git worktree from a ready pool, so parallel workers never touch each other's files; waiting on a worker costs the supervisor no tokens. | Node.js, no dependencies |
| sheepdog | Live kanban board and dispatcher for dozens of Claude Code and Codex sessions in herdr. Every session is a card filed by hand into Focus, Ongoing or Tools; the board raises one verdict, whether the next step is on you, and counts a session as busy only when a process, a counter or a timer proves it. Each card carries a one-line recap and a prompt-cache countdown; a dispatcher session reads the whole board as text and writes decisions back. | Node.js, no dependencies |
| local-dictation | Push-to-talk dictation that runs entirely on a Windows PC: faster-whisper on the GPU transcribes speech full of product names, and a local LM Studio model restores punctuation and terms. Every other change the model makes is diffed against the spoken words and rolled back, and every setting in config.toml carries the measurement that chose it. |
Python, faster-whisper, LM Studio |
| multi-lane-development | Delivery board that ran sprints on a fleet of parallel Codex agents. It handed each ticket to a free lane, had a separate agent check every pull request against the spec, reviewed the code, merged on green CI and walked the live site afterwards; I was asked once per sprint. | Node.js, no dependencies |
I am in Riga, Latvia. Write to me on LinkedIn.