← The AI Hype Audit — all 100 verdicts
HOLDS UP An open-source "agentic-first" CRM — the AI research agent IS the product, free and MIT-licensed
The claimA startup open-sourced the CRM it built for itself: a durable research agent that runs on its own schedule, keeps working after you close the browser, writes strong evidence straight to the record and hands weak evidence to a human — sandboxed, self-hosted, free.
The rare viral post that's basically all true — because it mostly just reads the project's own README, which undersells instead of overselling. The repo is real (MIT, launched July 31 by compliance startup Comp AI), and every architecture claim checks out against the code and docs: 18 authored tools plus 4 plain-English skill files, an explicit no-confidence-scores design, a sandbox with deny-all network egress and no database credentials, and a durable work queue so the agent schedules its own rechecks. The "20 stars" humility line was true at post time; it blew past 2,500 stars in three days. Even the warnings are honest: self-hosted, real setup (Bun, Docker, Google OAuth), promising-not-proven. One miss: the reel says "built on Claude" — the repo's default model is a different, configurable model; the reel creator added that detail, and built none of this himself.
What holds up
- Repo verified: trycompai/crm on GitHub — MIT license, created July 31, 2026, actively maintained, ~2,535 stars by Aug 3
- README matches the caption nearly word-for-word: 18 tools, 4 markdown skills, evidence-vs-suggestion split, sandboxed egress, durable queue
- Vercel's "eve" durable-agent framework is real (launched June 2026) and is exactly what the architecture leans on
- Comp AI is a real funded company (trycomp.ai, $2.6M pre-seed) that built this for its own sales team first
What doesn't
- "Built on Claude" is the reel's own embellishment — the default agent model in the code isn't Claude (it's swappable)
- The poster built nothing: it's a screen-record of the founder's launch post, riding a 24-hour star spike, with a comment-"CRM" DM gate for engagement
- Days old, ~zero production track record — the README says treat it as promising, not proven. Believe it.
The catch
Almost none on the product claim — the catch is the messenger. The reel account farms engagement off other people's launches, and the one detail it added ("built on Claude") is the one detail that's wrong. Read the source, not the reseller.
How to actually do it
- Skip the comment-"CRM" DM gate — the repo is public: search GitHub for the Comp AI open-source CRM
- Read the README's design sections first; the no-confidence-scores and sandbox reasoning are worth stealing even if you never deploy it
- Only self-host if you're comfortable with Docker + OAuth setup — this is a weekend project, not a signup button
- Treat it like any 3-day-old dependency: pilot on non-critical data until it has a track record
Logged as a genuine Build-It candidate: the durable-research-agent pattern (agent keeps its own notes, humans confirm weak evidence) rhymes with how we already run our own memory systems.
- Effort
- An afternoon to read and evaluate; a weekend to actually self-host
- Cost
- $0 software (MIT) + your server + model API usage
- Stack
- Bun, Next.js, Postgres, Docker, Google OAuth, a durable-agent framework, model of your choice
- Confidence
- High
- Posted by
- an AI-content account screen-recording another founder's launch post for comment-bait
We test hype for free. We build the real thing for a living.
Thirty minutes, no pitch — and you'll leave with something useful either way.
Book a call with Todd or start with the free Business Checkup →The Verdict Weekly
Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.
© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.