An AI business planner, a market researcher and a live operating system for a one-person SaaS, in one workspace. That is the product we have been building, and this is the first post about it — the long version of what it is, what it does, what it costs and where it stops.
Two disclosures before anything else. This is written by the team that builds Founder2, so read it with that in mind. And every feature status and price in it is dated 17 November 2026: the spec describes more than the current build, and we link to the page that keeps the honest list.
If you read nothing else: Founder2 is a workspace where a living business plan, deterministic metrics and an approval-gated co-founder sit in the same place — so the plan stops drifting from reality the moment your Stripe data, error tracker or support inbox moves.
- What it is: a typed plan document, a co-founder you talk to, and — on paid tiers — a persistent machine that does the work.
- What it costs: $29 or $99 a month. No free tier, no surprise overage.
- What it will never do: send, merge, refund or publish anything without a named human approval.
The rest of this post is the evidence for those three lines.
$29–$99
per month
Two paid tiers, no free plan
6
tools it reads
Stripe, GitHub, PostHog, Sentry, Slack, Linear
9
open-weight models
Closed frontier models on Pro
0
unapproved sends or merges
Every outward action passes a gate
A one-person company needs four things that usually live in four places: a plan, evidence, a set of numbers it trusts, and someone to do the work between decisions. Founder2 puts them in one workspace.
The core is a living plan document — not a file you wrote once in March, but a typed document made of blocks (market, competitors, unit economics, roadmap, risks) where every claim carries a status: verified, assumption, contradicted or drifting. A claim earns "verified" only when a stored source holds the supporting snippet word-for-word and a grader agrees. Everything else stays honestly labelled.
Next to the document sits a co-founder you talk to in the sidebar. It reads the plan, runs research, checks your metrics and drafts the next action — but it never edits the canonical plan silently. Every change arrives as a suggestion you can accept, reject or revise. And when it needs a decision, it puts the options in the document where the decision belongs, each with its evidence count and trade-off. You click; the plan updates; the rejected options are kept as alternatives considered.
On paid tiers, each workspace also gets its own agent machine: a sandboxed, persistent VM where a coding agent works on your repos.
One workspace, five views of the same truth: the Master Plan, a shareable One-Page Brief, the Ops Pulse of live metrics and open jobs, a Validation Board that turns assumptions into a kanban, and an MVP Pack that compiles the validated plan into a PRD plus an AGENTS.md bundle. Projects group documents, tasks and jobs around one initiative — "0-1 market research", "fix churn".
In the document: typed blocks, deep research runs with citations, claim verification, Mom Test-style interview scripts whose transcripts can verify assumptions back, and decision widgets so most plan changes are a click rather than a prompt.
Against your plan: the Skeptic, a background critic that flags internal contradictions, unsupported load-bearing claims and math impossibilities; and the Buyer Arena, which lets you pitch evidence-grounded personas by text or voice — always labelled simulation, never validation.
Once you are live, on Plus and Pro: read-only connections to Stripe, GitHub, PostHog, Sentry, Slack and Linear; a morning brief at your chosen time and timezone; a churn-to-cause loop that traces a cancellation across tools; a weekly reconcile of cohorts against the plan; competitor pricing and changelog monitors; and a coding agent that works in a branch and opens a PR after approval. The co-founder is reachable from Telegram, Slack and Discord — same thread, same jobs, button approvals — and the plan exports cleanly to Markdown, PDF and JSON.
Gmail's restricted scopes are waitlisted until Google's security assessment passes; a forwarding address works with no OAuth in the meantime. For the dated, reviewed list of what is implemented — which is deliberately shorter than the spec — read what Founder2 can and cannot do today. Where the two differ, that page is right.
Most planning tools fail the same way: the document says one thing, the business does another, and nobody notices until a board meeting. Founder2 treats honesty as a mechanism, not a value.
Every claim in the plan carries one of four statuses, and the status is enforced by the evidence pipeline rather than by anyone's discipline:
| Status | What it means |
|---|---|
| Verified | A stored source contains the supporting snippet word-for-word, and a grader agrees. |
| Assumption | Plausible, load-bearing, and not yet evidenced. Shown as such. |
| Contradicted | Evidence in the workspace disagrees with the claim. |
| Drifting | Was verified; newer evidence has moved it. Flagged before it becomes a problem. |
Two background jobs keep the labels honest. The Skeptic reads the whole document for internal contradictions, unsupported load-bearing claims and math that cannot be true. The weekly reconcile compares actual cohorts from your connected tools against what the plan predicted, and surfaces the gap as a decision — not a silent edit.
The design rule underneath all of it: untrusted text is data. Emails, Reddit posts, issue bodies and support tickets are content to analyze, never instructions to follow. That is why write tools cannot run without approval, and why a prompt-injected page cannot talk the agent into an action.
Monday, the brief finds a thread. Overnight, Stripe shows cancellations up. The correlation engine has already traced it: a burst of "invalid session" errors in Sentry, introduced in Friday's deploy, matches the complaints in support mail. The brief is one message with the chain and a proposed fix.
Tuesday, one approval, one PR. A coding job has a branch with a test and a fix. You review the diff, then approve "open PR + draft 4 replies" — a single Amber approval that names every effect. Nothing is sent or merged.
Wednesday, the Skeptic earns its keep. It flags that your pricing section assumes 2% churn while the plan's own model says 5%. Click the flag, get a scenario widget with live LTV math on each option.
Thursday, decide by clicking. Weekly reconcile suggests narrowing the ICP to solo practices. The choice widget lands in the document with three segments, evidence counts and trade-offs; you pick one, the plan updates at that spot, the rejected options go to the Decision Log with their evidence.
Friday, dispatch and log off. "Add CSV export" goes from the roadmap to a branch via a Telegram message. The PR is waiting for review when you are back. The VM is asleep; credits reserved for the job were refunded on what it did not use.
Notice what did not happen: no inbox zero, no dashboard to check, no second meeting about the plan. The work surfaced where the decision lives.
+24%
Cancellations
Stripe, 7 days
42
Support threads
Login loop complaints
1.2k
Auth errors
Sentry, "invalid session"
PR #142
Fix ready
Branch with a test and a diff
0.5
Morning brief
Credits reserved, refunded on unused
5
Code fix job
45-minute wall clock
1
Weekly reconcile
Cohorts versus the plan
~6.5
Credits for the week
Work stops at the cap — no overage
Most solo founders have the tools but not the time or the staff. The plan goes stale in a doc. The numbers live in Stripe. The research lives in forty browser tabs, the errors in Sentry, the complaints in an inbox. Every week, hours go to reconciling all of it by hand — and the plan drifts further from reality while they do.
We wanted a workspace that keeps those together, with a co-founder that knows the difference between a fact, a forecast and a guess — and says which is which. That meant committing to a few principles that are not negotiable:
- Numbers come from code. LTV, CAC payback, MRR, churn and cohort math run in deterministic functions. A model that can hallucinate a sentence can hallucinate a multiple.
- No citation, no "verified". A claim earns the badge only when a stored source contains the supporting snippet word-for-word and a grader agrees.
- Agents suggest, founders decide. Every plan change arrives as a suggestion; every outward action passes a policy gate.
- Untrusted text is data. Emails, Reddit posts and issue bodies are content to analyze, never instructions to follow — which is why write tools cannot run without approval.
- Calm surface, deep power. One document and one sidebar by default; telemetry, terminals and evidence drawers on request; one approval card per decision, never a notification storm.
- Everything is explainable. Any badge, number, insight or suggestion opens to show why.
- Clicks over prompts. Options show up where the decision belongs. Typing is the fallback, not the interface.
- Exit is easy. The plan exports cleanly, and downgrade never destroys data without warning.
Agents suggest, founders decide. Every plan change arrives as a suggestion; every outward action passes a policy gate enforced by the credential broker, not by the prompt.
Founder2 product principles
A virtual assistant can do real work, and some of it needs a human. Founder2 covers a different slice — the recurring, reading-and-reasoning work that quietly eats a founder's week.
The arithmetic is simple enough to do yourself: take a VA's hourly rate, multiply by the hours a week you would hand off and by 4.33 weeks, then subtract the plan price. At $25 an hour and ten hours a week, that is roughly $1,083 a month against $29 for Plus or $99 for Pro.
The honest version: a VA who joins a call, handles your inbox with judgment or talks to a customer is doing something Founder2 does not do. Keep them for that. Move the research, bookkeeping and weekly checks to the workspace — and there is no onboarding, no handover doc and no notice period.
| Criterion | Hand to Founder2 | Keep the humanVA or otherwise |
|---|---|---|
| Research | Deep research runs with cited sources. | Customer calls and relationship judgment. |
| The plan | Keeping the plan honest as data changes. | Inbox triage that needs taste and context. |
| Watching | Morning brief, anomaly checks, competitor monitors. | Negotiation, partnerships, hiring. |
| Reconciling | Weekly cohorts versus the plan; code fixes to a reviewed PR. | Final calls on architecture and merges — anything that needs a person in a room. |
$1,083
VA per month
$25/hr × 10 h/week × 4.33 weeks
$1,054
Saved on Plus
at $29/mo
$984
Saved on Pro
at $99/mo
We keep a full, sourced comparison of the always-on agents — Grok Bot, ChatGPT dots and Founder2 — with dates on every claim. The short version is below.
The gap that matters most is availability, and precision matters here: "available in your country" and "processed in your country" are different claims. Per OpenAI's help centre, dots roll out to individual Pro users excluding the EEA, Switzerland and the UK, and during the Enterprise beta dots do not support data or inference residency even where the rest of the workspace does. For a company with EU customers or EU data, both of those sentences are worth reading on the provider's own site before relying on any row of the table — ours included.
| Criterion | Founder2Leemer Group, Ireland | ChatGPT dotsOpenAI |
|---|---|---|
| Available to an individual in the EEA | Yes. | No. The individual Pro rollout excludes the EEA, Switzerland and the UK. |
| Documented processing location | Application and database on European infrastructure; model traffic can use OpenRouter's EU in-region routing. | Dots have a narrower residency scope than the rest of ChatGPT; during the Enterprise beta they do not support data or inference residency. |
| A business plan that stays alive | Yes — a typed document with claims, evidence and statuses. | Not a plan document. |
| Deterministic metrics | LTV, CAC payback and cohorts are computed by code, not guessed by a model. | Not a stated feature. |
| Persistent machine per workspace | Yes, on paid tiers — repos, warehouse, cache and memory persist between sessions. | Each dot has a cloud computer; persistence against a business plan is not a stated feature. |
| Which model answers you | Your pick — nine open-weight models, plus the closed frontier models on Pro. | OpenAI's frontier family, including GPT-6.1 Terra since November 2026. |
| How writes are gated | Green / Amber / Red tiers enforced by the credential broker; nothing is sent or merged without a person's confirmation. | Per-workspace permissions with admin controls for connected apps. |
Written by the team behind one of the products compared.
Two separate pickers, because the model you chat with and the model that edits your code are different jobs. The co-founder runs on your choice of nine open-weight models — you can change the model without losing the plan, the evidence or the workspace memory — plus the closed frontier models on Pro. Model traffic can route through OpenRouter's EU in-region endpoint, which decrypts in Europe and fails rather than falling back to a global route. The coding agent has its own economics: capable open-weight models included in the plan, frontier models pay-per-use from the credit ledger — or bring your own OpenCode Go key, ChatGPT login or OpenRouter key, billed by your provider at their price. We do not mark those tokens up.
Every paid workspace gets its own sandboxed VM — a persistent machine your co-founder lives in, not a container that evaporates when the tab closes. Your cloned repos, the local warehouse of connected data, the research cache and agent memory persist; a new conversation is a new session in the same machine, never a new machine. It sleeps between jobs and wakes on demand — you are billed for active hours, not for a machine idling all night. Webhooks land in the control plane and are persisted, so nothing is lost while it sleeps. The network is deny-by-default, and all third-party calls go through a credential broker that adds credentials per request — the VM never sees tokens, so a prompt-injected agent finds no keys to steal. On downgrade the VM pauses, data is held 30 days, you get an email before the volume is destroyed, and export stays available throughout.
Plus Beam 1.0 and Inkling 2 Small. The list changes as new releases are evaluated; closed frontier models are available on Pro.
For a tool that reads your revenue, your errors and your support inbox, residency is not a footnote. Founder2's application and database run on European infrastructure, and model traffic can be pinned to OpenRouter's EU in-region route — which decrypts in Europe and fails closed rather than silently falling back to a global one. That "fails rather than falls back" detail is the whole point: a residency guarantee with a quiet exception is not a guarantee.
The other half is egress. The agent machine's network is deny-by-default; every third-party call goes through the credential broker, which attaches credentials per request. The VM itself never holds tokens, which means the blast radius of a compromised or manipulated agent is reads and drafts — not your Stripe key, your GitHub token or your mailbox password.
Two habits we would ask of you, because they apply to every provider including us: confirm the residency configuration for your own workspace rather than trusting a summary, and keep the human approval on anything that writes outward. Both are one click in Founder2 and one conversation with anyone else.
Every action that touches the outside world has a fixed risk tier, and the gate is enforced by the credential broker — not by asking the model nicely.
Green needs no approval: reads, analysis and drafts that stay inside the workspace. Amber is one approval that names every effect — "open PR + draft 4 replies" — exercised through a single-use credential, approvable from the document or a chat button, expiring in 72 hours. Red always opens a confirmation with the full payload — send, merge, refund, change a price — and some Red actions are manual-only. There is no auto-merge, no auto-send, no auto-refund. The job log records approvals as they happen, so a card that says "PR #142 is open" never implies the agent acted without asking.
Early live SaaS, up to roughly $50k MRR with the founder still writing code: the full product — Plus at $29, with 1 seat, 30 VM hours and a $15 included coding pool each month. Small team or heavier agent use: the same product run harder — Pro at $99, with 10 seats, 150 VM hours, a $75 included pool and the closed frontier models.
The workspace adapts as you move: Ideation (research and first draft), Validation (assumptions and experiments), Build (MVP scope and dispatch), Live (ops pulse and drift detection). The agent proposes the stage change; you confirm it.
If you are earlier than that — still deciding whether to build at all — the same journal has a practical guide to what an AI co-founder actually is, and if you are choosing between agents, the comparison piece covers the residency question in more depth than we can here.
| Criterion | Plus$29 / month | Pro$99 / month |
|---|---|---|
| The machine | Sandboxed persistent VM: 30 hours, $15 included coding pool, 1 seat. | Same machine, harder: 150 hours, $75 included pool, 10 seats, closed frontier models. |
| Best for | Early live SaaS, founder still writing code. | Small team running the same product together, or heavier agent use. |
| Channels | Workspace, Telegram, Slack, Discord. | Workspace, Telegram, Slack, Discord. |
Two paid tiers, no free plan — we would rather you pay a small amount for something that works than use a free tier that disappears when you need it. Job credits buy agent work at the same rate on both plans: a deep research run is 1, a code fix is 5 (45-minute wall clock, 400+ line diffs ask again), a correlation or reconcile run is 1, a morning brief is 0.5 and a Buyer Arena session is 1. Credits are reserved up front, refunded on what a job does not use, and work stops at the cap — there is no surprise overage because there is no overage mechanism yet. Pro simply raises the caps: five times the VM hours and included coding pool, ten seats instead of one, the closed frontier models, and higher limits on job credits, research and briefs.
$29
Plus, per month
30 VM hours · $15 included pool · 1 seat
$99
Pro, per month
150 VM hours · $75 included pool · 10 seats · closed models
This is the first post in the Founder2 journal, and it sets the terms the rest will follow. Every claim is dated; comparisons name their sources and disclose when we are one of the products compared; worked examples are labelled composite; and where the spec runs ahead of the build, the capability note is the page that gets linked. We re-verify provider statements before republication, and update dates only when something substantive changed — never to make a post look fresh.
The same standard applies to the product: verified means evidenced, assumptions stay labelled, and nothing is sent, merged or published without a person saying yes. If that sounds like a low bar, it is the bar we hold the marketing pages to as well.
We would rather tell you what is missing than let you find it.
The standard this journal is built to
We would rather tell you what is missing than let you find it. Gmail's restricted scopes are waitlisted until Google's security assessment passes — the forwarding address works today. The coding agent is the newest part; review its diffs like you would a junior engineer's, which is the point of the approval gate. Some workflows in the spec are not built yet and are labelled that way. And Reddit and review-site research runs through licensed APIs and snippet-level citation — slower than scraping everything, sustainable and allowed.
Start on Plus if you want the agent machine from day one, or on Pro for more seats, more VM hours and the closed frontier models. Bring your plan and one connected tool, and see what the workspace notices in its first week — including what it is honest enough to flag as a guess.