{"slug": "ai-company-brain-how-to-build-one-in-5-steps", "title": "AI Company Brain: How to Build One in 5 Steps", "summary": "A developer built an AI-powered 'company brain' agent named Dojo for their startup, which automates daily standups, tracks progress, and generates reports. The agent runs on a Mac mini with a dedicated Hermes profile, uses Discord for communication, and relies on DeepSeek as a cost-efficient LLM. The entire kit is open-sourced, and the developer shared a five-step guide for building one.", "body_md": "**Quick answer:** An AI company brain is a dedicated agent instance that lives in your team's chat, runs daily standups, tracks progress toward your goals, and generates reports automatically. I built one for my startup, named it Dojo, and open-sourced the entire kit. This guide covers the architecture, the 5 setup steps, and how to run it for your own team.\n\nI first saw the idea in Antiwork's Gumclaw, the agent that now runs all of Gumroad's support and operations. It stuck with me: if an agent can run a company's day-to-day operations, it can at least run our standups. So I built our own version. This post is how it works, what it costs, and how you can have the whole thing free.\n\nBefore Dojo, our founding team ran operations the way most small teams do: standups happened in one channel when people remembered, GTM tracking lived in spreadsheets nobody updated, and weekly reports took an hour to compile and were read by no one.\n\n| Pain Point | What It Cost Us |\n|---|---|\n| Standups were voluntary | No daily rhythm, no accountability |\n| Metrics in spreadsheets | Numbers were always 2 weeks stale |\n| Reports took manual effort | Nobody had time to write them |\n| Targets felt abstract | No one could answer \"are we on track?\" |\n\nI wanted a system where the ops loop ran itself: the agent reminded, collected, compiled, reported, and tracked. Humans just answered and reviewed.\n\nA company brain is not a chatbot. It is an agent instance with a job: it knows your team, your targets, and your product; it lives in your chat; and it executes workflows on a schedule.\n\nThe architecture looks like this:\n\n```\nAlways-on machine (we use a Mac mini)\n  └── Hermes Agent (dedicated profile)\n        ├── Discord gateway (team channel)\n        ├── SOUL.md (identity, always loaded)\n        ├── Skills (loaded on demand)\n        │     ├── daily / weekly / monthly / quarterly workflows\n        │     ├── meeting notes, support triage, sales leads\n        │     ├── project tracking, onboarding\n        │     └── roadmap, team, company knowledge, self-learning\n        ├── Cron jobs (automated triggers)\n        └── Data folder → JSON files → Dashboard HTML\n```\n\nThe key design choice: everything is files and schedules, not magic. Daily standups compile into JSON, weekly reviews read those JSON files and generate a dashboard, and monthly reports read the weekly files. The whole system is inspectable and debuggable.\n\nYou need a machine that stays on. A Mac mini, a spare laptop, or a small server all work. We used a Mac mini already sitting in the office.\n\nInstall the agent and create a dedicated profile for the brain. Giving it its own profile keeps its memory, skills, and config separate from anything else on the machine.\n\nPick a cost-efficient LLM as the provider. We run ours on DeepSeek, which keeps the monthly API bill low. The brain is not doing heavy reasoning — it is parsing standups and compiling reports — so you do not need a frontier model.\n\n\"The brain is not a genius. It is a reliable operator. Cheap models are enough when the workflows are well-defined.\" - Me, after the first month of running Dojo\n\nThe brain lives where the team already talks. We use Discord.\n\nThe setup takes about 15 minutes: create a bot in the Discord Developer Portal, generate an invite URL with the right permissions, and add it to your server. The permissions matter more than people expect. The brain needs to read messages, send messages, read history, and mention people. Without mention permissions, your standup reminders are silent.\n\nCreate three channels so the brain has clear lanes:\n\n| Channel | Purpose |\n|---|---|\n`#standup` |\nDaily standup input and reminders |\n`#reports` |\nWeekly, monthly, quarterly reports |\n`#alerts` |\nNags, escalations, error alerts |\n\nChannel structure is how the brain learns where things happen. Consistency matters more than the specific names.\n\nSkills are the brain's playbooks. Each one defines a workflow: triggers, steps, parsing rules, and output format. Our kit ships 13 skills across three categories:\n\n| Category | Skills | What They Do |\n|---|---|---|\nMeetings |\ndaily, weekly, monthly, quarterly workflows + meeting notes | The full cadence: standup to strategic review |\nOperations |\nsupport triage, sales leads, project tracking, onboarding | The day-to-day work beyond meetings |\nKnowledge |\nroadmap plan, team, company/product, self-learning | What the brain knows, kept current |\n\nThe self-learning skill is the sleeper feature. It watches conversations for signals like \"X joined the team\" or \"project Y shipped\" and updates the team and project files automatically, with confirmation. The brain maintains itself instead of rotting.\n\nCron jobs drive everything. The daily standup fires on weekdays, the weekly review on Friday afternoon, the monthly report on the 1st, and the quarterly review on the 1st of January, April, July, and October.\n\nTime zones matter. All schedules run on your team's local time, so the standup reminder lands when people are actually at their desks. This is the kind of detail that separates a system people use from a system people ignore.\n\nThe brain needs a place to store what it learns, and a baseline to compare against. Create the data folder structure, then add a baseline JSON with your team, your current numbers, and your targets.\n\nThis baseline is what makes the reports meaningful. \"We added 120 leads this week\" is a number. \"We added 120 leads against a target of 140\" is a signal. The whole reporting loop is built on that comparison.\n\nHere is the rhythm once everything is running:\n\nFor our GTM tracking, the metrics are demos booked, leads added, and proposals sent. Yours will be different — that is why the kit keeps metrics as placeholders. The system works for any numbers you track.\n\n**Quick question:** \"What if nobody answers the standup?\"\n\n**Answer:** The brain nags politely, then escalates. Three days of silence from a project triggers an alert to leadership. The system does not force participation, but it makes silence visible.\n\nThe weekly workflow generates a self-contained HTML dashboard with charts: targets versus actuals, per-person contributions, trend lines, and a filter bar. It is generated from the JSON files, so it is always current.\n\nWe expose ours through a tunnel so the whole team can check it from anywhere. The dashboard is the answer to \"are we on track?\" that used to require digging through spreadsheets.\n\nThis is the part people ask about most. Our numbers:\n\n| Item | Cost |\n|---|---|\n| Machine | Mac mini we already owned (a used one is ~$300-500) |\n| LLM API (DeepSeek) |\n$2-5/month for our usage |\n| Discord | Free |\n| Total ongoing | ~$2-5/month |\n\nThe brain runs standups, weekly reviews, monthly reports, and GTM tracking for less than the cost of one team coffee run. That is the whole pitch.\n\n**Challenge: The brain misses replies or parses them wrong.**\n\nExample: A teammate writes \"yesterday I did demos\" without numbers, and the brain logs zero.\n\nFix: The parsing rules are explicit and documented. If a metric is ambiguous, the brain asks \"how many?\" rather than guessing. Corrections also feed the self-learning loop, so the same mistake rarely happens twice.\n\n**Challenge: People stop replying to standups after two weeks.**\n\nExample: Participation drops from 100% to 60% by week three.\n\nFix: This is a process problem, not a tool problem. The brain's nags help, but the real fix is leadership modeling: founders answer first, every day. A standup system only survives if the people at the top treat it as non-negotiable.\n\n**Challenge: The reports feel repetitive.**\n\nExample: The weekly report says the same thing three weeks in a row.\n\nFix: That is information, not a bug. If nothing changed, the team needs to ask why. The quarterly workflow adds strategy and hiring checks precisely so the cadence escalates in depth as the period grows.\n\n**Challenge: The brain goes quiet and nobody notices.**\n\nExample: The gateway drops and standups stop for a day.\n\nFix: The alerts channel catches this. Add a health check so the brain itself is monitored. An ops brain that silently dies is worse than none.\n\nNo. Any always-on machine works: a spare laptop, a small server, or a Raspberry Pi class device. The brain is not compute-heavy. What matters is that the machine stays on and connected.\n\nAbout 2-3 hours the first time: machine setup, Discord bot, skills, cron jobs, and a dry run. The kit's README walks through every step. A second install takes under an hour.\n\nThe agent works with Discord, Slack, Telegram, and others. We run ours on Discord. The kit's commands and workflows are written for Discord but adapt easily.\n\nThe brain stores data as local JSON files on your own machine. Nothing is sent to a third-party SaaS. You control the machine, the data, and who has access. The LLM API calls carry only the specific prompt text.\n\nYes. We run on DeepSeek for $2-5/month. The workflows are well-defined, so a cost-efficient model performs fine. You can upgrade to a stronger model later if needed.\n\nThe kit uses placeholders for metrics, targets, and team members. You replace them with whatever your team tracks: demos, sales calls, support tickets, engineering velocity, anything.\n\nA chatbot answers questions. A company brain executes workflows on a schedule: it collects standups, compiles reports, generates dashboards, escalates blockers, and updates its own knowledge. The difference is the loop, not the chat.\n\nA company brain turns team operations into a running system: standups that actually happen, targets that are actually tracked, and reports that compile themselves. We built ours for $2-5 a month, and the entire kit is open source on GitHub. Start with the daily standup workflow, run it for two weeks, and add the weekly and monthly layers once the rhythm sticks. The direction of the industry is clear — YC just open-sourced QM, a multiplayer harness for agent work. Company brains are becoming a category, and you can build yours this weekend.\n\n**Next step:** clone the [Company Brain Kit](https://github.com/sanjuacodez/company-brain-kit), follow the 10-step README, and run your first standup tomorrow. If you are building agent-driven teams, see how we think about the org chart model at visibility.so.\n\n*Found this useful? Share it with a founder who still compiles standups by hand. Follow @sanjayshankarr for more on running teams with AI agents.*", "url": "https://wpnews.pro/news/ai-company-brain-how-to-build-one-in-5-steps", "canonical_source": "https://dev.to/sanjay_shankar_5bc7676af0/ai-company-brain-how-to-build-one-in-5-steps-4lcm", "published_at": "2026-08-12 07:56:07+00:00", "updated_at": "2026-08-12 08:16:08.848111+00:00", "lang": "en", "topics": ["ai-agents", "ai-products", "developer-tools", "large-language-models"], "entities": ["Dojo", "Hermes", "Discord", "DeepSeek", "Mac mini", "Antiwork", "Gumclaw", "Gumroad"], "alternates": {"html": "https://wpnews.pro/news/ai-company-brain-how-to-build-one-in-5-steps", "markdown": "https://wpnews.pro/news/ai-company-brain-how-to-build-one-in-5-steps.md", "text": "https://wpnews.pro/news/ai-company-brain-how-to-build-one-in-5-steps.txt", "jsonld": "https://wpnews.pro/news/ai-company-brain-how-to-build-one-in-5-steps.jsonld"}}