Torno is the agent-native system of record for growth: one Loop from signal to shipped
change to honest readout. Agents do the operating; you govern what ships, spends, and
publishes — and every turn comes back knowing more.
A readout carries its bounds, basis, and freshness. “No data”, “below k”, and “not yet sufficient” are answers — never rendered as zero.
Agents operate, humans govern
API, CLI, and MCP are generated from the same capability map as the UI. Anything irreversible, sensitive, or spending passes through policy and approval.
Never in your critical path
Assignment runs locally from signed, cached bundles and fails open to control. No Torno network call sits in your customer's request path.
Batteries included
Ingest, listening, analytics, experiments, revenue, ads, briefs, governance — the instruments the Loop needs, in one account.
Listen
Every observation gets a front door.
Product events, support email, public threads, review feeds, Metric watches — each lands as an immutable Signal carrying provenance, consent, and trust. Nothing the world tells you gets silently lost.
SIG-799behaviorPlan-picker step exit rate rose to 61%funnel · owned events owned
SIG-812voiceTwo customers asked which plan includes custom domainssupport inbox owned
SIG-816voice“Pricing page didn't tell me what the Studio plan is for”reddit · public web public· quarantined
SIG-820metric Δtrial_to_paid_28d moved −0.4pt week over weekMetric watch owned
Signals are immutable — corrections arrive as new Signals. Public-web content stays
fenced as data and everything derived from it keeps the untrusted-derived label.
Diagnose
Analysis that shows its work.
Versioned Analyzers read the evidence — language models only ever see fenced, quarantined text — and produce Insights that carry confidence, citations, and freshness. Claims inherit the trust of what they came from.
Trials stall at plan selection; plan-difference questions dominate recent voice.
SIG-799SIG-812SIG-816untrusted-derived2h ago
Propose
Proposals that cite their evidence.
A Hypothesis fills a forced grammar: because this evidence, we believe this change, for this audience, will move this Metric, by this much, within this window. No field is optional — vibes don't compile.
collapsing three plans to two with a comparison row
for
new trials
will move
trial_to_paid_28d
by
+0.5pt
within
21 days
growth-agent
“Exit rate and the voice cluster point at plan confusion; a two-plan picker is the lowest-effort probe.”
reasoning required · stored as a claim
Ship
Ship with a measurability contract.
A Change declares its exposure — tag, goal event, assignment — or is honestly unmeasurable and excluded from causal claims. Policy holds anything irreversible for a human; delivery is signed, sticky, local, and fail-open.
CHG-31page_variantTwo-plan picker with comparison row
exposure tag
plan-picker
goal event
subscription_started
assignment
50 / 50 · sticky
first exposure
confirmed — measurement started
assignment bundle signed · synced 41s ago50% · ramp on evidence
change.activate → allow (scope: change_write)
APR-6ads.publish · Google campaign draft awaiting a human
human approval — shipped default for ad publishing
CHG-28Maker stories collectionno exposure declared — excluded from causal claims
Measure
Readouts that can't overclaim.
Sequential evidence with explicit states: trending is not a win, a guardrail breach can auto-pause, and below-k stays suppressed. Stripe-derived revenue reads through the same semantic layer as everything else.
−0.1 … +0.9pt1,842 exposuressample-ratio check · passingcheckout_completion · no breach
not yet sufficient — the readout says so
Trials · last 28dn = 214
Studio-plan trials · NLbelow k (k = 25) · suppressed
The Needlebottega · prod
MRR$8,420+$310 this month
leading pair: trials_started / wk
new+$540
expansion+$120
contraction−$150
churn−$200
live Changes aimed at this target
CHG-31 trending
CHG-28 insufficient evidence unmeasurable
Decide
Decisions become memory.
Ship, kill, iterate, or extend — recorded with rationale and evidence snapshots in an append-only log. The next proposal starts from everything this turn learned. That is what returning means.
Corrections are new records, not edits. The next proposal reads this log —
torno context makes proposing without it impossible.
webbottega · prod
LP-4Lift trial → paid conversion measuring
targettrial_to_paid_28d 3.4% · 28d rolling
liveCHG-31 · 1,842 exposures trending
pendingAPR-6 · ads.publish · Google campaign draft
last decisionDEC-12 ship · Annual-first pricing page (CHG-24)
torno context LP-4
$ torno context LP-4
statemeasuring · turn 6
targettrial_to_paid_28d 3.4% · 28d · owned events
liveCHG-31 · 1,842 exposures · trending
guardrailscheckout_completion ok · activation_rate ok
pendingAPR-6 ads.publish · awaiting a human
last decisionDEC-12 ship · annual-first pricing
Agent-native
Your agents get the whole product.
Every registered capability generates its web, API, CLI, and MCP bindings from one
map — an agent is never boxed into a reduced surface. Mutations require reasoning,
which is stored as a claim and never treated as instruction. And when an action
needs a human, the agent waits at the approval — it does not route around it.
Growth work usually dies in the gaps between six tools. Torno bundles the
instruments the Loop needs — collection, listening, analytics, experiments,
revenue, ads, briefs, governance — behind one semantic layer, so every number
means the same thing everywhere it appears.
The boundary is deliberate: payment rails and ad auctions stay with their
providers. Torno bundles what the Loop needs — not everything with a dashboard.
Briefsstructured handoff to humans or coding agents
Policyscopes, approvals, spend ceilings, audit
Agent surfacesAPI, CLI, MCP — one capability map
Status
Where Torno is today, honestly.
Torno's whole premise is that numbers should never say more than their evidence.
The same rule applies to this website.
ImplementedEvery product vertical — Loops, ingest, identity, analytics, experiments, revenue, ads, briefs, policy, agent surfaces — has production code behind it.
Locally tested1,496 behavior scenarios covered by local test suites, 38 explicitly exempted, 0 unmapped — snapshot of 2026-07-22.
In progressLive provider verification: Stripe, Google and Meta ads, listening connectors, imports, WorkOS sign-in, deployed SDK middleware. Each connection shows its verification state in the product.
OpeningEarly access, in small turns. Join the list and we'll return — that's rather the point.
Who it's for
Built for small teams with real stakes.
Founders and solo builders
Growth work that never happens because it lives across six tabs. One account, one Loop, agents doing the legwork — and you approving the parts that matter.
Growth and product teams
Seed to Series B, where every experiment must count. Shared Metric definitions, pre-registered evidence, and a decision log the whole team compounds on.
Coding and marketing agents
First-class principals with scopes, budgets, and required reasoning. The same capabilities as the UI — through API, CLI, and MCP — with policy holding the line.
torno— Italian: “I return.” From tornare, to turn on a lathe.
Start the first turn.
Bring a Property, wire the SDK, define one Metric — and let the Loop start
listening. Every turn after that comes back with more.