Projects

Each of these is in production, or was until it was retired.

Acquirelytics

Acquirelytics

Finds businesses for sale across more than 800 sources and uses AI to screen each one against a buyer’s thesis. Brad founded it in 2024 and still directs it.

acquirelytics.com

432,717listing links collected
65,186AI verdicts against a buyer’s thesis
953broker CIMs read and written up by AI
3,191merged PRs, whole team
335MCP tools

Acquirelytics: 432,717 listing links in, 953 CIMs read by AI

Mar 2025 to 18 Oct 2026. AI verdicts are counted from Sep 2025, and a listing judged for two buyers counts twice. Bars are on a log scale.

About 250,000 listings in 20 months

From 807 configured sources: broker sites scraped in bulk, dedicated marketplace scrapers, a Chrome extension, broker websites, and broker email read by AI. Source: Acquirelytics desk metrics, Mar 2025 to 17 Oct 2026.

What it does

  • AI cleans each listing, matches its industry and merges duplicates.
  • A buyer’s plain-English thesis becomes gates and deal-breakers, and every verdict shows which check a deal failed.
  • A waterfall shows where deals drop out, and what loosening one threshold would add.
  • A pipeline from first contact to NDA to the broker’s deal memo, with broker email and follow-ups in one place.
  • AI reads each broker’s CIM and writes a deal memo, a one-pager and a valuation.
  • Buyer profiles and a managed tier for customer organizations.
  • 335 MCP tools, so the whole desk can be run from Claude.

The hard part

The verdict model changed three times in a month. Each candidate was tested first against people’s own pursue-or-pass calls on the same deals. Claude Haiku went live on 16 Oct because it agreed with the human call 81–90% of the time, for less than the model before it.

What’s different about it

  • Every model is trialled against human decisions before it goes live. Claude Haiku agreed with the human call 81–90% of the time, at about half a cent a verdict.
  • About 35 AI steps, each on its own model, usually for single-digit dollars a day.
  • Six scheduled Claude routines watch for errors, silent failures and AI spend, and 120 tools are off-limits to them.

Source: Acquirelytics admin tables and desk metrics, 18 Oct 2026.

Brain

The app Brad runs his AI agents from. Every Claude, Codex and ChatGPT session sits in one queue, ranked by what it needs from a person, next to the tickets, inbox and specs it relates to.

523merged PRs since 9 Sep 2026
45agent sessions live on 2 Macs
7,757agent sessions recorded
119MCP tools
209steps in the PR gate

Why it exists

Brad runs about 50 agent sessions at once, across two Macs and several tools. At any moment some are stuck on a question, some need approval and some are finished and ready to test, and nothing said which. Brain does: every session reports in when it starts and joins one queue, and each can be answered, approved, stopped or started from any device.

What it does

  • Every agent session registers the moment it starts and lands in a queue ranked by what it needs from a person: stuck, a question, an approval, ready to test.
  • Answer, approve, stop or start a session on any Mac running agents, from a desk or a phone.
  • Move a session to another Mac, or from Codex to Claude, with its history and work in progress.
  • Projects, boards and tickets. Sessions are filed to tickets automatically.
  • Finds earlier sessions that touched the same work.
  • One inbox for Gmail, iMessage, WhatsApp and Slack, with a Priority view and filing to tickets.
  • Agents publish pages: specs, reviews and small tools, with per-version approval.
  • AI usage history: every token across coding agents and billed API keys, by project, model and month.

The hard part

There’s no staging: merging is deploying, with about 50 agents committing. So every incident becomes a check that runs on every change, and the house rules are thirteen practices, each written after the incident that taught it.

What’s different about it

  • Every agent’s work gets its own database and preview link, so any change can be reviewed anywhere, phone included.
  • Anything that starts or steers an agent needs a tap on a hardware key.
  • Mostly built by agents, with a person reading every change before it merges.

Source: Brain and GitHub, 18 Oct 2026.

Sir Nigel

Sir Nigel

Sir Nigel was Brad’s OpenClaw instance: an open-source agent framework he ran around the clock on a Mac Studio for eight months, as a chief of staff. It was the first version of Brain.

openclaw.ai

7,680agent sessions
29services on its private network
15+models, cloud and local

What it did

  • Ran the launch QA for Dad Cards: 24 browser checks, which caught and fixed an admin sign-in bypass. It also built the joke pipeline, with four models writing candidates and a fifth judging.
  • Built the controller for the Vestaboard at home, with every write read back and checked, and schedule changes sent as a diff to approve.
  • Built an Alexa white-noise skill, and when two Echos started failing it found the error in the logs and shipped a fix the same morning.
  • Did research overnight: mapped 669 businesses in one regional market from 210 searches, and screened 624 software businesses for sale down to five serious candidates.

Why it was retired

Brad retired it on 8 October 2026. It moved too fast, broke too often, and needed its own on-premises cloud to keep it going. Brain replaced it.

Smaller things

Side projects, each built mostly by AI agents. Some became small businesses.

112 billion tokens since November 2024

Two sources, each counted once: the coding agents Brad builds with (Claude Code, Codex, Cowork and OpenClaw), and the AI calls the apps make, as each provider billed them.

112Btokens, all sources
91.4Bcoding agents
32.6Bin September 2026
20.5Btokens in apps’ own API calls

Coding agentsApps’ own API calls

Most coding-agent tokens are the model re-reading its working context, which is normal for long agent sessions. Coding agents are counted from February 2026, when Brain started recording them; app calls from November 2024. Source: Brain’s /ai/history and the providers’ own usage reports, 18 Oct 2026.

Get in touch

Building something similar, or want to work with us? Email [email protected]. For consulting in the UK, see FireClaw.