Your startup runs on unproven assumptions. Find out which ones are true before they cost you.
Capture the beliefs behind your plan. Weigh the real evidence for each one. Decide with receipts. Work in the app, or let Claude and Cursor keep the ledger from your meeting transcripts.
No card needed. AI agents connect on every plan.
What we can prove today
- 13
- MCP tools
- From ingest_transcript to decide_assumption.
- 6
- Levels of evidence
- What customers did counts far more than what they said.
- 0–100
- Confidence, computed
- Moves only when real evidence lands.
- 0
- Scores an agent can write
- Every score comes from the database, not the model.
Works with Claude Code, Claude Desktop, and Cursor. Free plan included.
The loop
Capture, weigh, confront, decide
Four moves, run continuously. Your agent can drive all of them, or you can work the queue by hand. Same ledger, same scoring.
It enters the ledger while it is still a belief
A transcript pasted into the app, a claim typed straight into the queue, or your agent handing over the meeting it just sat through. Each one arrives as a triage item with a category, a domain, and an importance. Not a bullet in a doc nobody reopens. Triage is keyboard-driven: j/k to move, 1–5 to set priority, e to log evidence, . to decide.
Evidence is graded by what people did, not what they said
Every piece of evidence lands on a six-rung ladder, from repeated behavior at the top down to opinion and analogy at the bottom. Ten enthusiastic quotes are still worth less than one renewal, and the ledger says so out loud.
The dashboard is built to make you uncomfortable
Risk is importance × max(uncertainty, contradiction), so a belief people keep arguing with stays at the top instead of averaging out into safety. Evidence against you is not evidence you are safe. Below the risks, a category × domain cross-tab names the whole columns you have never questioned.
Record the call, keep the receipts
Each dossier plots risk and confidence per snapshot, so you can see whether a belief is firming up or quietly drifting. An assumption becomes decision-ready when enough real evidence lands, or earlier when it is decisively contradicted. Decide it, and the reasoning stays on record forever, including for your agent.
The wedge
One command, and the agent you already have starts keeping the ledger
Assumption Mapper ships a published MCP server with thirteen tools. Add it to Claude Code, Claude Desktop or Cursor and your agent can capture assumptions from a meeting, grade the evidence, plan a test, and report what actually moved.
$ claude mcp add assumption-mapper \
--env ASSUMPTION_MAPPER_API_KEY=am_your_key \
-- npx -y assumption-mapper-mcpCreate a key under Settings → Account → API keys, paste it in, restart the client. Claude Desktop and Cursor take the same server as a JSON block; the exact config lives on the connect page.
Claude Code — assessment brief, excerpt
You
Here’s the transcript from this morning’s Delphi call. Log it against NovaKit.
Claude
→ ingest_transcript(projectId: novakit, extraction: 2 assumptions, 2 evidence)
Ingested delphi-2026-07-24.txt · 1 source · 2 evidence items
Assessment:
Evidence logged: 2 items · 1 independent source
Strength mix: stated intent ×2
Assumptions moved (1):
• confidence 50 → 59 · risk 4.00 → 3.20 · 2 new (2 supports)
Usage-based pricing beats per-seat for expansion revenue
Drift flags (1) — stated as settled fact, still undecided:
• "EU data residency is table stakes for mid-market" —
named as a requirement, not as a hypothesis.
Cautions:
⚠ All 2 new items trace back to a single source — that is
one independent data point, not 2.
⚠ All 2 new items are stated intent — what people say they
would do, not what they did.Both signals are stated intent from one call, so I’ve left the pricing assumption in triage rather than moving it to active. If you want to test it, give me the kill criterion first — “we abandon this if … by …” — I can’t log an experiment without one.
Works on every plan
Your agent extracts with its own model, so agent-side ingestion carries no plan gate. Server-side extraction — pasting a transcript into the app — is the part that needs Pro or Team.
Your tokens, not ours
On the agent-side path our server runs no AI call at all. It validates, stores, scores, and hands back the brief. The only model that reads your transcript is the one you already pay for.
Scores stay server-side
The write schema rejects risk, confidence and evidence mass outright, and rejects unknown keys with a 400 rather than dropping them quietly. Your agent submits raw evidence; Postgres does the arithmetic.
Why it holds up
Three things a doc cannot do for you
A Notion page will hold any belief you put in it, at any confidence you claim, for as long as you like. That is the whole problem.
Evidence that can’t flatter you
Supporting and contradicting evidence are counted separately, and stronger evidence counts for more. A belief with six warm quotes and two hard contradictions does not average out to fine. It stays loud until you decide.
“Your agent will just fill it with AI slop.”
It can’t. An agent submits raw evidence: a quote, a direction, a strength, a source. Every score is computed on our side, and there is no field an agent can use to talk itself into a better number.
Evidence strength
Weight
- Repeated behavior1.00
Renewals, recurring usage or purchases.
- Committed behavior0.80
Pre-order, signed LOI, paid pilot, deposit.
- Engagement0.60
Return visits, referrals, unsolicited requests.
- Stated intent0.30
“I would use it / pay for it / switch.”
- Opinion0.15
Expert, advisor or team opinion. Secondhand reports.
- Analogy0.05
Reasoning by analogy, with no external data.
Ten enthusiastic quotes weigh about the same as one renewal, one signed pilot, and one referral.
Kill criteria, pre-committed
An experiment cannot be recorded without a written kill criterion: “we abandon this if <observable> by <date>”. It is a NOT NULL column, a required field on the API, and an instruction the agent is told not to satisfy on your behalf. Closing the experiment does not decide the assumption; it only makes it ready to decide.
“We already do this in Notion.”
Notion never refuses. It will happily hold an experiment with no exit condition, and a criterion agreed after you have seen the result is not a criterion. This one asks you first and will not proceed without it.
Your blind spots, named out loud
Four categories (desirability, viability, feasibility, adaptability) crossed with ten domains, from ICP to pricing to competition. A single empty cell in a 44-cell grid means nothing. An empty column means you have never questioned pricing at all, and the dashboard says so in as many words: Never questioned: Pricing, Team, Competition.
“We know what we haven’t looked at.”
Nobody does. The gap is invisible precisely because nothing in it ever comes up. That is what makes it a blind spot rather than a backlog item.
Where your data sits
- Row-level security, on every path
- The API enforces the same RLS as the web app: an API key is exchanged for a short-lived, RLS-scoped token, so it sees exactly what its owner sees. A key cannot mint, list or revoke keys.
- Encrypted in transit and at rest
- TLS on the wire, encryption at rest in the database. Operated from the EU by TheMotionBridge; every subprocessor is named in the privacy policy.
- Your transcript can stay with your model
- When your agent extracts, our server makes no AI call. It validates, stores and scores. Server-side extraction is the one path that sends text to Anthropic, our AI subprocessor.
- No ad trackers, no data sale
- We run no third-party advertising trackers and we do not sell personal data. Delete your account and your data is removed within 30 days.
Full detail in the privacy policy.
Pricing
The agent works on the free plan
Paid tiers buy volume and seats, not the method. The strength ladder, honest scoring, kill criteria, and coverage are on every plan.
Free
Get started. No credit card needed.
- 30 active assumptions
- 1 project · 3 team members
- MCP server and agent-side ingestion
- Evidence logging and strength grading
- Risk scoring, coverage and staleness
- Decision log and export
Pro
Most usefulFor builders running multiple ventures.
- Everything in Free
- Unlimited assumptions
- 20 projects
- In-app AI extraction from transcripts
- Export and reporting
- Priority support
Team
For venture studios and innovation teams.
- Everything in Pro
- 20 team members
- Shared assumption library
- Admin dashboard
The free cap counts active assumptions, so deciding one frees a slot.
Questions
The ones that actually come up
What counts as evidence?
Do you train on my data?
Which agents work with it?
Can my team use it without an agent?
What happens to my transcripts?
Can the agent inflate its own scores?
Name the beliefs your business is betting on. Then let something other than optimism grade them.
Start with one project and the assumptions you already argue about. Connect your agent when you want the meetings to file themselves.
€0 forever · 30 assumptions · MCP on every plan