# Nix Inc > Nix Inc sets up Claude Code and Claude Cowork for small teams, writes the rules the agents work under, and reviews what they produce. Clients buy their own Claude Team seats. Nix Inc is an AI consulting practice founded by Chris Nixon. Contact: chris@nixinc.co. The site is one page at https://nixinc.co/ with these sections. Nix Inc sets up Claude Code and Claude Cowork for small businesses, writes the rules the agents work under, and reviews what they produce before it is used. Your team buys its own Claude Team seats. Founder Chris Nixon has spent fifteen years in digital analytics, paid media and SEO, up to director of digital analytics. ## 01 / What Nix Inc does What Nix Inc does for a team. Your company buys its own Claude Team seats and owns the accounts. Nix Inc does the work below on those seats. - Setup: Claude Code and Claude Cowork, configured on your company's Claude Team seats. - Rules: A written rule set for each agent: what it may do on its own, what it must ask a person about first, and what only a person does. The rules Nix Inc works under are on this page. - Review: What the agents produce is checked against the task and a written standard before your team relies on it. Changes near sign-in, payments or production get a deeper review. - First step: Email chris@nixinc.co. Say what your business does and one job you want an agent to take on. ## 02 / Builds Nix Inc's own builds. Nix Inc's own projects, not client work, numbered in the order each one started. Stopped and unfinished builds are listed too. Each card says where it stands, and each has its own page with how it works and what is not done yet. ### Build 01: NameIO (AI search visibility, live) Measures how often AI assistants recommend a brand. A free tool that checks whether AI assistants recommend a brand when shoppers ask buying questions. A full scan asks 20 prompts on Claude, ChatGPT, Perplexity and Gemini, scores all 80 answers and drafts fixes for the gaps. - Stack: Next.js, Firebase, Playwright, Claude Code CLI, Vercel, Zod - Details: https://nixinc.co/builds/nameio - Link: https://getnameio.com ### Build 02: Video pipeline (Automated video, stopped) Scripted explainer videos with local voice and captions. An experiment in making short explainer videos from a script with no hosted services: Remotion for the frames, a local text-to-speech voice and Whisper captions. It produced a test render and a 48-second EV comparison, then stopped. - Stack: Remotion, React, Piper TTS, Whisper, FFmpeg, Python - Details: https://nixinc.co/builds/video-pipeline ### Build 03: Mission Control (Agent dashboard, stopped) A first dashboard over local agent tooling. A Next.js dashboard that read Claude Code's local project folders and system stats to show sessions, skills, MCP servers and scheduled tasks in one place. Built in about two days, then replaced by the Hermes setup. - Stack: Next.js, TypeScript - Details: https://nixinc.co/builds/mission-control ### Build 04: Four-agent team (Multi-agent experiment, stopped) Scheduled role agents running a small business, unsupervised. An experiment in letting four scheduled AI agents, in CEO, CTO, marketing and research roles, run a small business on the open-source Paperclip orchestrator. They built an MCP server, a working SaaS prototype and a landing page, then stalled. - Stack: Paperclip (open source), Claude via CLI, TypeScript, Zod, Next.js, Prisma - Details: https://nixinc.co/builds/agent-team ### Build 05: Gemma 4 fine-tune (Model fine-tuning, stopped) One QLoRA run on past agent conversations. A pipeline that turned past agent conversations into training pairs and fine-tuned Google's Gemma 4 small model with QLoRA on a free Colab GPU. It ran once and produced a merged model, which was never evaluated or deployed. - Stack: Python, Unsloth, QLoRA, Google Colab, GGUF, Ollama - Details: https://nixinc.co/builds/gemma-fine-tune ### Build 06: Hermes Agent (Operations agent, running) A self-hosted operations agent that runs around the clock. A self-hosted deployment of Nous Research's open-source Hermes Agent. It takes instructions over Telegram, runs scheduled research jobs, and switches to a backup model when the main one is rate-limited or unreachable. - Stack: Hermes Agent (open source), Claude Code CLI, Claude Sonnet 5.5, macOS launchd, Telegram, SQLite - Details: https://nixinc.co/builds/hermes-agent ### Build 07: Daily market-intel scan (Scheduled research, running) A morning research brief and a weekly decision memo. Two scheduled Hermes Agent jobs. Each morning one gathers signals from Hacker News, Reddit, Google Trends and X, scores up to three ideas against written rules and files a cited note. Each Monday the other turns the week into one decision memo. - Stack: Hermes Agent scheduler, Claude Sonnet 5.5, Python, Hacker News API, Reddit RSS, Google Trends RSS, Telegram - Details: https://nixinc.co/builds/intel-scan ### Build 08: Agent-built knowledge base (Knowledge base, stopped) A wiki that an agent compiles from raw notes. An Obsidian wiki that an agent compiles from raw documents, based on an approach Andrej Karpathy described, with scripts to ingest, compile, search and check it. It reached 43 pages before it went dormant. - Stack: Python, Obsidian, gbrain (open source) - Details: https://nixinc.co/builds/knowledge-base ### Build 09: TrimIndex (Search pages from public data, live) Trim-level scores and buyer briefs for used EVs. A reference site that scores 1,134 trims of electric cars from model years 2020 to 2026, using NHTSA, EPA, pricing and owner data. In the next release, a written buyer brief publishes only when every number in it traces to a stored source. - Stack: Astro, React, TypeScript, Turso, Cloudflare Pages, Firecrawl, Claude Code CLI - Details: https://nixinc.co/builds/trimindex ### Build 10: domain-hunter (Internal research tool, paused) Finds expired domains linked from trusted sites. A command-line tool with a local web interface that finds expired or available domains linked from sites already trusted in a niche. It checks each domain's history, has Claude rate its old topic, rejects spam patterns and ranks the rest. - Stack: TypeScript, cheerio, DNS, whois, RDAP, Wayback Machine API, Claude Code CLI - Details: https://nixinc.co/builds/domain-hunter ### Build 11: Maillard (iOS app · Search over video, in development) BBQ cook plans cited to the second of source video. A native iPhone app that turns 2,995 transcribed cooking videos, mostly from BBQ creators, into a cook plan for the user's own grill and cut. Each step cites its source clip to the second, and fixed food-safety rules check the temperatures and holding times. - Stack: Swift 6, SwiftUI, FastAPI, Postgres + pgvector, Whisper, Ollama embeddings - Details: https://nixinc.co/builds/maillard ### Build 12: Claude Code setup (Rules and guardrails, in use) The rules, guardrails and skills Nix Inc works under. Nix Inc's own Claude Code configuration: a written working method, permission tiers, a deny list and a guard hook for the riskiest commands, a standard set of files in every product repo, and three custom skills. - Stack: Claude Code, CLAUDE.md, Skills, settings.json permissions, Bash and Python hook - Details: https://nixinc.co/builds/claude-code-setup ### Build 13: MeatBrains (iOS app · Camera reading, paused) Reads meat-case shelf tags and ranks packs by value. An iPhone app prototype that reads supermarket meat-case shelf tags with the camera, checks the arithmetic on each tag, ranks the packs in view by value and links recipes. A marbling feature was withheld because it missed its accuracy bar. - Stack: SwiftUI, Apple Vision, Swift Package Manager, XCUITest, FastAPI, SQLite - Details: https://nixinc.co/builds/meatbrains ### Build 14: Qwen 3.8 Splash, served locally (Private local model, running) A 27B model on the Mac itself, with no per-token bill. A Hermes Agent profile that runs on Qwen3.8-27B served on the Mac itself, with no cloud API and no per-token cost. Before the switch on 2026-09-19, a same-prompt benchmark showed about twice the speed of the previous local runtime. - Stack: Inco Splash 1.0, Qwen3.8-27B (4-bit), Hermes Agent, macOS launchd, Python, MLX - Details: https://nixinc.co/builds/local-model-splash ### Build 15: Local image pipeline (Local image generation, stopped) Text-to-image and edits on the Mac, from one command. A local text-to-image and image-edit setup on ComfyUI with the Qwen-Image 2.1 model, driven by a one-line shell command and documented in a runbook. The model weights have since been removed, so it does not run today. - Stack: ComfyUI, Qwen-Image 2.1, Bash - Details: https://nixinc.co/builds/image-pipeline ### Build 16: nixinc.co (Site, live) This site. Nix Inc's own site: a home page and one page per build, generated from typed content files, with a plain-text copy at /llms.txt, an RSS feed and a sitemap. Live since 2026-10-08. - Stack: Next.js, TypeScript, Vercel, Claude Code - Details: https://nixinc.co/builds/nixinc-co - Build 17: Next build: not chosen yet. It goes in the log before work starts. ## 03 / Founder Chris Nixon. Founder, Nix Inc. Fifteen years in digital marketing and analytics. Chris started in paid search at agencies, ran paid search for a Fortune 500 brand, then moved in-house to lead paid media, SEO, ecommerce analytics and site experience, up to director of digital analytics. Most of that work is measurement and governance: deciding what gets tracked, who is allowed to change it, how a change is tested before it goes live, and how the result is reported. He moved eight brands onto one shared ecommerce and content platform and owned the tagging standards for all of them. Nix Inc applies that approach to coding agents. Each agent works under written limits on what it may change, and its work is checked before anything ships. - Experience: 15 years, up to director level - Background: Digital analytics, paid media, SEO - Specialty: Measurement, testing and governance ## 04 / Who does what Who does what. The founder decides and signs off. Claude Code does the build work and Hermes Agent runs the scheduled jobs. - Founder: Chris Nixon. Writes the briefs, reads every change before it ships, and owns deploys, customer data and billing outright. - Builder: Claude Code. Works each ticket on a feature branch, inside the permission tiers listed under Rules. Nothing it writes ships until a person has read it. - Operations: Hermes Agent. Self-hosted and always on. Takes instructions over Telegram and runs the scheduled jobs, including the morning research brief. - Review: Claude Code, review pass. Checks each change against the written standard and marks it must fix, should fix or okay to ship. - Client teams buy their own Claude Team seats. - Deploys, customer data and billing are decided by people, not agents. - Not affiliated with Anthropic or any model vendor. ## 05 / Rules Four rules for every piece of work. They apply whether a person or an agent is doing the work. These are the four rules and three tiers Nix Inc works under. ### 01 Brief Nothing starts without a short written plan: which files will change, the smallest clean way to do it, what could go wrong, how it will be checked, and what is deliberately being left out of the first version. Plain terms: say what you are going to do, on one page, before you do it. ### 02 Ticket One task, one finish line, one change small enough to read in a diff. “Make it better” is not a ticket. If there is no clear finish line, the work does not start. Plain terms: cut the work small enough that someone can check all of it. ### 03 Eyes Use the finished thing the way a new customer would. Try the empty state, the error state and the success state. Open the database and check the record is actually there. Plain terms: click through it yourself before calling it done. ### 04 Review Three readers. The founder reads the change against the ticket. An agent reads it against the written standard and marks it must fix, should fix or okay to ship. Anything near sign-in, payments or production also gets a deeper automated code review before it moves. Plain terms: every change is read by a person and by an agent before it ships, and risky changes get a deeper check. ### What an agent may do on its own Three tiers, written into the instructions file Claude Code reads before it starts. The riskiest commands are also on a deny list in Claude Code's settings, which still applies when permission prompts are turned off: recursive deletes, force-pushes, opening a secrets file and writes to the payment account. - Go ahead: Read and inspect code; Propose a plan; Run the local tests; Edit on a feature branch; Update docs, draft a pull request. Plain terms: work that does not reach customers until a person approves it. - Ask first: Install a dependency; Migrate a database; Touch sign-in; Change anything about payments; Delete files or rename at scale. Plain terms: changes to sign-in, payments, the database or the software the project depends on. A person says yes first. - Only a person: Deploy to production; Decide anything about customer data; Decide anything about billing; Change anything security-sensitive; Open a secrets file. Plain terms: putting anything live, and any decision about customer data, billing or security. - Same rules for people and agents. - Hard blocks on the riskiest commands, in Claude Code's settings. ## 06 / Log The log. Decisions and releases, newest first, with the date and who decided. If one of Nix Inc's own builds is stopped, the reason goes here. - 2026-10-08 · Decision · Stopped and unfinished builds are listed too, each with how far it got: the video pipeline, Mission Control, the four-agent team, the Gemma fine-tune, the knowledge base and the local image pipeline. (Founder) - 2026-10-08 · Decision · Each build gets its own page, and five more builds are listed: the market-intel scan, domain-hunter, the Claude Code setup, MeatBrains and the local Qwen 3.8 Splash model. (Founder) - 2026-10-08 · Decision · The page now opens with what Nix Inc does for a client team and ends with a way to get in touch. It is written for small-business owners as well as developers, which replaces the 3 October decision. (Founder) - 2026-10-08 · Shipped · nixinc.co goes live. (Founder) - 2026-10-08 · Decision · Earlier builds go on the site after all: NameIO, Hermes Agent, TrimIndex and Maillard. This reverses the 2 October decision. (Founder) - 2026-10-03 · Decision · The site is written for developers first. Each rule also gets a one-line plain-English version for people who don't write code. (Founder) - 2026-10-02 · Decision · Focus: helping small teams run coding agents. Only new builds go on the site; earlier experiments stay off it. (Founder) - 2026-10-02 · Shipped · Research brief for this site: 32 comparable practitioners checked against their own pages, 477 sources. (Agents, checked by founder) - 2026-10-02 · Decision · Name: Nix Inc. Domain: nixinc.co. (Founder) - 2026-08-18 · Decision · Twice-daily research scans paused. (Founder) - 2026-06-17 · Note · nixinc.co registered. (Founder) ## Contact Write a few lines on what your business does, how many people would use Claude, and the job you want an agent to take on. - Email: chris@nixinc.co - Seats: Your own Claude Team plan - RSS: nixinc.co/feed.xml - Plain text: nixinc.co/llms.txt ## Links - Home: https://nixinc.co/ - Email: chris@nixinc.co - RSS: https://nixinc.co/feed.xml - Sitemap: https://nixinc.co/sitemap.xml AI consulting for small teams. Built with Claude Code. Not affiliated with Anthropic.