MyGentic Free Harness: high-level spec and PRD
A free entry point to MyGentic: 7 days of access and up to $0.50 of inference that MyGentic covers. It's built to bring whole communities (webinars, challenges, IR) onto the platform without hurting margin. Source: MyGentic Founders Product Sync (Wes, Nick, Aatu, Ronnie). Items marked OPEN are not decided yet.
$0.50inference budget per user (MyGentic absorbs it)
7 daysaccess window
Localingest on the user's machine (near-zero cost to us)
Whichever firstbudget or time runs out, then the upgrade prompt
1. Problem and goal
Not having a free way in is the biggest blocker to user numbers (Nick). People have to wait or pay before they try anything, and partners have no free platform to bring their communities into.
Goal: give people a free trial that shows the "aha" moment (their own knowledge graph plus a real conversation with their context) and then converts them to Authority. MyGentic's cost stays capped at about $0.50 per free user.
2. Who it's for
- Webinar, challenge and community audiences (IR / Kane), with no class purchase needed
- Solo creators and consultants evaluating MyGentic
- People who live in Claude or ChatGPT and want MyGentic over MCP. Wes flagged this as a large market the free tier still has to serve.
3. Decisions from the call
| Decision | Source |
|---|---|
| Target about $0.50 of absorbed inference cost per free user, plus 1 week of usage | Nick, 47:04 |
| Ingest and graph building run locally in the harness on the user's compute, using a small local embedding model. Inference does NOT run locally, because local LLMs aren't good enough. | Ronnie 22:50 / 26:55; Nick 42:03 |
| Use the cheapest model that still gives a quality result. Start with GPT-4.1 mini (≈1¢ per recall loop, so ≈50 questions). Also test cheaper models. | Ronnie 40:09; Nick 41:02, 43:20; Wes 42:59 |
| Don't restrict connectors. Let users build a big, impressive graph; the limit is on how much they can engage with it. | Nick, 47:04 |
| Show the full UI. Paid features are visible but locked behind "needs Authority" | Nick, 42:03 |
| Don't power the free tier with users' own ChatGPT/Claude subscriptions, because of the account-ban / ToS risk. MyGentic pays for the free inference. | Nick 36:23; Aatu 37:15 |
| Build it, then test models with Ashish's existing testing harness (the same one used for the Qwen tests) | Nick, 43:20 |
| Not available right away. It's the next big product project ("a couple of weeks"). Wes owns the free-tier feature spec, and Nick owns Authority/Genius/Legacy for the site. | Nick 58:09, 58:24; Wes 59:22 |
| Keep the naming simple. Aatu proposed calling mini-brains just "brains", and Wes agreed. | Aatu / Wes, 50:17 |
4. Scope
| In (free) | Out / locked (upgrade) |
|---|---|
| Install the harness locally; local ingest, chunking and embedding All connectors; personal knowledge graph / Atlas view Chat with own context through MyGentic-hosted inference, up to $0.50 Visible map of every paid feature | Usage beyond $0.50 or 7 days Brain building with tests/evals (Wes: token-hungry, likely not free) OPEN Paid brains / Moat purchases MCP use from Claude/ChatGPT OPEN |
5. $0.50 budget mechanics (proposed)
- Server-side metering. All free inference goes through a MyGentic proxy that records provider token usage × model price for each call. We don't trust a client-side count, because Ronnie noted the local harness can't count tokens.
- Display. Users see a simple meter ("questions left ≈ N") rather than dollars. OPEN
- Soft warnings at 50% and 80% used. Hard stop at $0.50: chat is disabled, the graph and ingest stay viewable, and an upgrade card appears.
- Pre-flight check: reject a call if its estimated cost is more than what's left, so we never overshoot.
- The model is set by a server flag, so we can swap 4.1 mini for a cheaper model after testing without a release.
6. 7-day lifecycle (proposed)
- Day 0. Sign up (email verified), install the harness, connect sources, and see the graph come alive. The clock starts at first sign-in. OPEN: or at install?
- Day 1–6. Nudges: "ask your brain X", show what's left, and preview locked features.
- Day 5 and Day 7. Reminder emails with the upgrade path.
- End (day 7 or $0.50, whichever comes first). Inference is paused. The user keeps their local graph, and one click to Authority keeps everything going.
- No auto-renewal of the free allowance. Nick also raised "one month free vs. rolling" (33:14), and the call settled on 7 days.
7. Abuse and cost guardrails
- One free allowance per verified email, device and payment-free account. Detect duplicate sign-ups.
- Rate limit per minute and per day so nobody can burn the budget in one burst.
- Promo stacking: free status is recorded in the promotions system (the Promotions Hub Aatu is building), so a free user can't stack it with other promos, and partner-funded extensions get billed to the partner.
- A global daily spend cap on the free pool, with alerting.
8. Success metrics (proposals, not facts)
- Activation: % who connect ≥1 source and ask ≥1 question
- Free→paid conversion by day 14
- Real average cost per free user vs. the $0.50 target
- Answer-quality score for the free model vs. the paid model (Ashish's harness)
9. Rough phases
- Model test. Run 4.1 mini plus cheaper models through Ashish's harness, and measure cost per loop and quality.
- Metering proxy + entitlements. Build the $0.50 / 7-day state and the promo-system flag.
- Harness free mode. Local ingest, locked-feature UI and upgrade flow.
- Pilot with one community (e.g. an IR webinar), then the public site tier.
10. Risks
- A cheap model makes the product look weak, and people don't upgrade (Wes, 29:31)
- Free is good enough that nobody pays (Nick, 25:13)
- Cost per query is still an estimate (~1¢/loop, from Ronnie's recollection)
- Concept sprawl: brains, mini-brains, skills, plugins (Aatu / Nick)
- Course brains change over time and need cloud hosting, not static local files (Wes, 45:36)
11. Open questions for Aatu, Nick and Ronnie
- Do free users get any brains (e.g. an IR trial brain), or none? (Nick, 47:04)
- How does the MCP / Claude-ChatGPT market get a free path if free inference only runs through our harness?
- Is brain building included? If so, what's the limit?
- Does the clock start at sign-up, install or first query? Is there a grace period?
- Exact model after testing: 4.1 mini or cheaper? And the confirmed cost per recall loop.
- Can a partner (IR) fund extended free time, and how is that billed?
- Should the meter show questions or dollars to the user?
- Windows/Mac harness readiness for free users, and support load.
- Final naming: "brains" only?
Draft by Wes for discussion. Based on the Fathom recording of the MyGentic Founders Product Sync, 27 Oct 2026. Nothing here is final until agreed.