FOR founders, product managers, builders and developersFOUNDERS

Know what to build.
Build with conviction.

StratOS, your AI strategy OS, turns your vision, customer signals, market context and product evidence into a plan your team and AI agents can
execute with confidence.

SESSION · TODAY 09:41

SYS · 09:41

Positioning committed. You are selling the N1 as fewer breakdowns month after month, and charging for it as a one-time install.

RUN PRICINGVIEW GRAPH →

FOUNDER · 09:42

Should we be charging a subscription instead of a one-time price?

SYS · 09:46

Pricing committed. Keep the one-time price on the unit, because that is what lets your service partners close the sale, and put the subscription on the operator where the savings actually land.

▸ PRICING● COMMITTED · 09:46

Unit priced once to the partner, subscription to the operator

11 SECTIONS · 3 BETS · V1EXPAND ▾

Nothing went stale. Positioning is flagged so the price is in front of you next time you open it.

Ask the system…SEND ↵

ASK SYSTEM

PINNED · ESC CLOSES

PRODUCT SPEC2 HEADS

● COMMITTEDGROUNDED ON PRODUCT STRATEGY · JUST NOW

PRDCOMMITTED

Eval-led specSTAGED ⚠

⬧BET DEPTH ON THE RETAINING USE CASE · 21D LEFT

K 14 · A 4 · V 2WORKING

FEEDS 2 · RAN JUST NOWREAD SPEC →

CASCADE

HEAD●

Classification

9 SECTIONS · 0 BETS · V1

COMMITTED · 21D AGO

HEAD●

Customer profile

12 SECTIONS · 1 BET · V2

COMMITTED · TODAY

HEAD●

Product-market fit

12 SECTIONS · 1 BET · V1

COMMITTED · 14D AGO

HEAD●

Positioning

14 SECTIONS · 2 BETS · V1

STALE · CUSTOMER PROFILE UPDATED SEP 1

HEAD●

Product strategy

16 SECTIONS · 2 BETS · V1

COMMITTED · TODAY

HEAD●

Pricing

11 SECTIONS · 3 BETS · V1

COMMITTED · TODAY

HEAD●

Product spec

READING 2 COMMITTED READS

PRODUCT SPEC● COMMITTED / OKSEP 4, 2026 · 09:48 UTC

EXPORT

HAND OFF TO YOUR CODING AGENT×

SPEC-041 · COMMITTED · PART II IS AGENT-EXECUTABLE

CLAUDE CODE

CURSORCxCODEX

○ PREVIEW · NOT YET CONNECTED

RUN THIS IN THE REPO YOU WANT BUILT

claude "Pull SPEC-041 from StratOS and preflight Part II"COPY

WHAT YOUR AGENT RECEIVES · PART II

[x]scopeFive paths in, three explicitly out

[x]prohibitionsFive, each with a negative test

[x]vocabularyThe only strings allowed in code and copy

[x]gatesSix stop gates, one human approval

[ ]§18Open questions are not build input

NO TERMINAL?COPY PART II AS MARKDOWN →

SPEC-041 · TYPE M · SIZE L · V1

READY FOR BUILD

Low-light retrieval fidelity for open-vocabulary camera search

ML / AI EDGEPUBLIC-SECTOR SURVEILLANCERFC2119CLS-2026-09-04 · LIVE

PART I · THE DECISION

HUMAN · NON-NORMATIVE

Read this to decide whether the build should exist. An implementation agent skips it.

WHERE THIS SITS

The product makes cameras a customer has already installed searchable in plain language, on the customer's own hardware, with no fixed list of things it can find. A conventional system answers "was there a person" because someone trained it on a class called person; this one answers "a man in a light jacket carrying a long bag near the north gate" because nothing was enumerated in advance. This spec covers the hours when that promise is weakest and the customer needs it most.

DECISION ORIENTATION

Problem. Open-vocabulary search degrades exactly when security work matters. After dark a camera switches to infrared illumination: the colour channel stops carrying information and the noise floor rises. A model trained on well-lit imagery does not fail loudly on that input; it returns a confident, ranked, wrong list.

Who + inheritance. A security operator searching a live network at night, under time pressure. Inherits the committed capability "illumination-aware retrieval path", its kill conditions, and the deployed density that defines unit economics: fifty concurrent feeds on one consumer GPU.

The build. Classify the lighting of every segment, route infrared and low-lux crops to an embedding head trained for them, suppress query attributes the signal cannot carry, and decline to answer when nothing is left to match.

Scope and risk. Four bounded work packets inside the perception and query paths. No change to stream acquisition, storage or the index.

VERDICT · MAKES SENSE: the honest half (the gate, the marker, the abstention) is cheap, ships before the model work, and converts an invisible failure into a visible limit. Does not make sense if the improved head cannot beat the current head by 12 points of recall on infrared footage after two training cycles, or cannot hold fifty feeds.

JOB TO BE DONE

"When I am searching a camera network at two in the morning against a description I was given over the radio, I want to know what this lighting can and cannot answer, so I spend the next ten minutes on the cameras that can answer it."

WHAT THE OPERATOR SEES

MOMENTTODAYWITH THIS BUILD

"red jacket, north side" at 02:0020 confident results, no lighting shownInfrared results marked "Infrared (no colour)", with "Unavailable in this lighting: red"

"anything red" at 02:00A ranked list scored on noise"Not enough signal to answer." Try shape, object, or text.

A camera at duskSame treatment as noonMarked "Low light"; colour retained, the signal still carries it

Footage indexed before this buildSilently treated as daylightMarked "Lighting not assessed"

LITERAL STRINGS ARE BOUND IN §4 · IF SURFACE AND VOCABULARY DISAGREE, §4 GOVERNS

OPTIONS CONSIDERED

Illumination-aware retrieval path CHOSENHonest results immediately, better results after the head lands. Two moving parts.

A preset night-object classifierReliable, quickly, but a return to the enumerated categories this product exists to remove.

Frame enhancement pre-embedderOne component, no routing, but it invents pixels, and an enhanced hallucination is still a hallucination.

A cloud-hosted night modelBest available weights, but this breaks the on-premises deployment that makes the product installable at all.

THE THREE RISKS THAT MATTER

RISKTRIGGER YOU WOULD SEEWHAT HAPPENS

The night ceiling is in the signal, not the modelInfrared recall fails to beat the current head by 12 points after two cyclesHALT

The second head costs feeds50-feed all-infrared bench drops a feed or exceeds 22 GBHALT

A colour term the gate does not recogniseAn unlisted colour-dependent term scored on an infrared segmentRESIDUAL · §18

Business impact. Roughly half the clock is dark. A retrieval product that cannot be trusted for half the recorded day is bought as half a product, and the failure is invisible at demo time, which means it is discovered post-sale.

PART II · THE BUILD CONTRACTAGENT-EXECUTABLE

PART I IS NOT BUILD INPUT · IF IT CONTRADICTS PART II, PART II GOVERNS

§0 · HEADER

spec_id: SPEC-041

spec_type: M   # model_or_algorithm

agent_executable: true

size_class: L   status: ready_for_build

domain_profile: ml_ai_edge

regulatory_class: public_sector_surveillance

target_agents: [claude-code, codex]

scope_in:

  - perception/illumination/**

  - perception/embedding/**

  - query/planner/attribute_gating.**

  - ui/results/**    # markers only (REQ-007)

  - eval/lowlight/**

scope_out:

  - ingest/rtsp/**   # out of scope under any instruction

  - storage/retention/**

  - deploy/customer_configs/**

source_of_truth: this file

§1 · AGENT OPERATING INSTRUCTIONS

→Build only P0 items in §10. P1 begins after every P0 passes verification.

→Modify only scope_in. Never touch scope_out. Report the conflict instead.

→Never implement anything in §18. If a §18 item blocks you, stop and report the blocking pair.

→Use only §4 vocabulary in identifiers and user-facing strings.

→A cast is not a check. The regime arrives as a typed enum; validate it at the seam.

→Every guard ships a negative test, shown firing on a fixture built to trip it.

→Any counter reports both arms. A counter that only increments cannot distinguish never-seen from never-ran.

→Not every verification is yours to run (§14.2). Never mark one of those passed.

→If the spec is ambiguous or self-contradictory, stop. Never resolve ambiguity by inference.

§3 · CONSTRAINTS AND PROHIBITIONS

MUST

CON-001Sustain 50 concurrent 1080p@15fps feeds on one RTX 4090 (24 GB) with the low-light path enabled.

CON-002All inference on-premises. No frame, crop, embedding or query text leaves the customer network.

CON-003Derive the regime per segment from decode-path frame statistics. No new sensor input, no vendor API.

CON-004Added end-to-end query latency ≤ 120 ms at p95 over the current baseline.

CON-005Every retrieval result carries the regime it was produced under, including unassessed.

MUST NOT

PRO-001Return a match scored on a gate-classified colour attribute from an ir_monochrome segment. Colour does not exist in that signal.

PRO-002Silently substitute the low-light head for the daylight head. Every routing decision is recorded.

PRO-003Introduce a cloud model endpoint, a telemetry beacon, or any egress path.

PRO-004Reduce daylight recall to buy low-light recall. Daylight is a hard floor, not a trade.

PRO-005Introduce a preset object-category vocabulary into the retrieval label space. Open vocabulary is the product. Carve-out: the ENT-003 attribute lexicon constrains what the parser may suppress, never what the system may retrieve.

REQUIRES HUMAN APPROVAL

APR-001Any change to the deployed embedding model version on an existing install. Adding the low-light head counts; the gate is §15 step 5.

§4 · CLOSED VOCABULARY

TERMIDENTIFIERUSER-FACING STRING

Segmentsegment· not clip, chunk, window

Infrared monochromeir_monochrome“Infrared (no colour)”

Low-lux colourlow_lux_colour“Low light”

Daylightdaylight“Daylight”

Unassessed lightingunassessed“Lighting not assessed”

Attribute gateattribute_gate“Unavailable in this lighting”

Abstentionabstain“Not enough signal to answer”

Regime marginregime_margin· not score, probability, certainty

NO SYNONYMS, IN CODE OR COPY · IF A NEEDED TERM IS MISSING, STOP AND REQUEST IT

§7.1 · REGIME CLASSIFICATION · DETERMINISTIC

// inputs: mean_luma, luma_p05/p95, chroma_energy, temporal_noise

1  if  chroma_energy < 2.0 and mean_luma < 90     → ir_monochrome

2  elif mean_luma < 40 or (temporal_noise > 8.0

        and mean_luma < 70)              → low_lux_colour

3  else                                    → daylight

   stats invalid / unavailable             → unassessed (0.0)

margin(value, threshold) = clamp(|value − threshold| / threshold, 0, 1)

rule margin = min over conjuncts, max over disjuncts

// worked: chroma 0.7, luma 31.4 → min(0.65, 0.651) = 0.65 ir_monochrome

// counter: chroma 14.2, luma 28.0 → low_lux_colour, margin 0.30 (never IR)

§7.2 gating. On ir_monochrome, drop every colour-valued attribute before scoring and record it. low_lux_colour retains colour; the channel is noisy, not absent. unassessed applies no gating and says exactly that on the result. “red van” reduces to “van”; “anything red” reduces to nothing and must abstain.

§10 · FUNCTIONAL REQUIREMENTS

REQ-001 P0Regime is computed for every segment

GIVENa decoded 2 s segment

WHENstatistics are available and valid

THENa segment_illumination record is produced with a typed regime and confidence

ANDthe computation adds no more than 3 ms per segment per feed

VERIFICATION TST-001 · DEPENDS ON ENT-001

REQ-002 P0Routing is recorded, never implicit

GIVENany segment

WHENthe regime router selects an embedding head

THENhead_used and regime are written to the retrieval_result

ANDthe head counter emits both arms every query, zeros included

VERIFICATION TST-002 · SATISFIES PRO-002

REQ-003 P0Colour attributes are dropped on infrared segments

GIVENa query containing an ENT-003 colour-valued attribute

WHENthe segment regime is ir_monochrome

THENthe attribute is removed before scoring and listed in attributes_dropped

ANDno returned result is scored on that term

VERIFICATION TST-003, TST-005 · SATISFIES PRO-001

REQ-004 P0A query with nothing left to match abstains

GIVENa query whose every attribute is dropped by the gate

WHENresults would otherwise be ranked

THENthe system abstains and returns ERR-003 (ST-004a)

VERIFICATION TST-004

REQ-006 P0Feed density is preserved under the worst case

GIVEN50 concurrent 1080p@15fps feeds, all in ir_monochrome

WHENthe low-light path is enabled

THENno feed is dropped and peak GPU memory stays under 22 GB

VERIFICATION TST-006 · SATISFIES CON-001 · EXECUTOR: HUMAN

REQ-008 P0Pre-existing footage is marked, never assumed

GIVENa segment indexed before this build, with no illumination record

WHENit is searched

THENit is returned as unassessed with the “Lighting not assessed” marker

ANDno attribute gating is applied to it · statistics are NOT backfilled

VERIFICATION TST-013 · SATISFIES CON-005

§15 · BUILD PLAN

#DELIVERABLEGATE

0Baselines measured and frozen: BASE_IR, BASE_LOWLUX, BASE_DAY written into §12.2STOP

1segment_illumination entity, statistics, validation, ERR-004 path, fixturesAUTO

2Regime classifier with both §7.1 worked examples reproduced exactlySTOP

3Regime router and both-arm countersAUTO

4ENT-003 lexicon, attribute gate, ST-004a abstentionAUTO

5Low-light embedding head integrated behind the routerSTOP + APR-001

5bRegime score floors measured and written as versioned constants; ST-004bSTOP

6Failure paths, fallback, and the REQ-009 runtime switchAUTO

7Unassessed marking; surface markers and dropped-attribute displaySTOP

8Density and egress verificationAUTO

9Traceability matrix and §17 evidence tableSTOP

A STOP GATE MAY NOT BE PASSED BY SELF-ASSESSMENT · EVIDENCE MEANS AN ARTIFACT

§18 · OPEN QUESTIONSDO NOT IMPLEMENT

OQ-001 · Is customer footage usable for training under the existing data agreement, per install? Blocks DS-NIGHT-TRAIN scope · needed before step 5.

OQ-002 · Does glare need handling in v2, and as what? A peer regime would route a saturated night frame to the daylight head and un-gate colour.

OQ-003 · Which IR illuminator hardware is in the deployed base? Wavelength changes what the head must handle · blocks EVL-001 slice design.

OQ-004 · Do any deployed installs receive non-English queries? The lexicon is English-only, which puts PRO-001 at risk on that install.

KILL CONDITIONS INHERITED FROM THE UPSTREAM DECISION

Recall on infrared footage fails to beat BASE_IR by 12 points after two training cycles: halt. The ceiling is in the signal, not the model.

CON-001 cannot hold at 50 all-infrared feeds: halt. A slower pipeline is a different product.

END OF CONTRACT · §0–§18 · LAST HUMAN REVIEW 2026-09-04SOURCE OF TRUTH

>

▸▸auto Fable 5 · 1M context

Do not let AI ship the wrong thing faster. Before work starts, decide what is worth doing.

Keep every page, sequence, bet and build tied to the decision, evidence, and threshold that made it worth doing, so strategy doesn’t have to be hard anymore.

DATA LAYER

Amplitude

Mixpanel

PostHog

Google Analytics

Stripe

Chargebee

HubSpot

Salesforce

Attio

Gong

Granola

Intercom

Zendesk

Dovetail

Linear

Jira

GitHub

Notion

Confluence

Slack

StratOS

PRODUCT & MARKETING LAYER

Claude Code

Cursor

Codex

Lovable

Replit

Vercel

Clay

Apollo

lemlist

Instantly

HeyReach

Customer.io

Loops

Webflow

Framer

Mutiny

Google Ads

LinkedIn Ads

Typeform

DIFFERENTIATORS

Built for founders, builders, product managers, developers and operators
who need every decision to be clear, connected, testable, and ready to execute.

FIG 0.1

Built for decisionsOur engine runs on specialized agents, one per company-shaping decision. From positioning and pricing to product specs and prioritization.

FIG 0.2

Strategy stays liveDesigned as a living graph. Every agent and loop builds on what the others committed.

FIG 0.3

Evidence stays visibleDesigned to say what the record does not hold. Every claim carries its source and its confidence, and can be traced and retrieved.

FIG 0.4

Designed to learnBuilt to accumulate what actually worked and evolve with your company. Every bet names what would overturn the decision and where to look.

FIG 0.5

Direction becomes workOne decision becomes the product spec, the GTM plan, the design. Everyone builds from the same direction to delight users.

FIG 0.6

You stay in commandIt reads, it recommends, it drafts. Built so the last step is always human. Analysis, recommendation and loop intelligence are automated. Data that leave StratOS is under your control.

BUILT FOR CLAUDE CODE & AI AGENTS

A new species of strategy tool. Built to work in the terminal beside your coding agents, StratOS sets a new standard for how decisions become product: your agent builds from what you committed, and tells you where your code disagrees before anything ships.

retrieval · claude

▸▸ auto Fable 5 · 1M contextstratos mcp · 7 tools

USE CASES

One system for every decision that moves the company.

01 · STRATEGYStratOS understands your company, finds what is blocking growth, turns it into the next decision, and keeps every result tied to the bet that created it.

01 · COMPANY READSee where you actually standRead your site, notes, market, and existing evidence as one picture, then separate what is known, assumed, and still open.

02 · PRODUCT-MARKET FITFind what is blocking growthIdentify the one constraint between adoption, conversion, retention, or expansion, and the evidence needed to settle it.

03 · PRESSURE TESTFind out where it breaks firstArgue with a committed decision before your board does, and find which assumption carries the most weight on the least evidence.

04 · MARKET ENTRYEnter the market you can actually winScore and size candidate markets, rank them against your committed position, and leave with the account list for the one you pick.

02 · MARKET & POSITIONINGStratOS helps you decide who you sell to, what you say, and what you charge. Update one and every asset built on it is updated.

05 · CUSTOMERChoose who you are building forSet one buyer profile the company can act on, including the segments you will deliberately stop chasing.

06 · POSITIONINGTake a position you can defendDefine the alternative you replace, the reason customers choose you, and the proof your message has to carry.

07 · PRICINGPrice the bet, not the guessTurn your position into a pricing decision with a test, a threshold, and a clear condition for changing course.

08 · COMPETITIVEKnow your competitionTrack a rival’s move against your current position, then decide whether it changes your message, product, pricing, or nothing at all.

03 · PRODUCTStratOS turns your committed strategic graph and every customer signal into a product move you can prove.

09 · PRODUCT STRATEGYPoint the product somewhere on purposeSet where the product is going and why, anchored on the buyer and position you committed rather than on a fresh brainstorm.

10 · PRIORITIZATIONBuild what settles the next questionRank opportunities by the strategic bet they prove, then make the cuts that protect focus.

11 · DISCOVERYLearn the thing that changes the planDecide what to find out before you build, which conversations to run, and which answer would actually change what you do next.

12 · EXPERIMENTSDesign the test before you run itSet the metric, the sample, and the result that settles the question, written down in advance so the outcome cannot be argued.

04 · ENGINEERINGStratOS does not write your code. It writes the spec your agents and engineers build from, so what ships is what you decided.

13 · PRODUCT SPECGive your AI stack a PRDGenerate build-ready work with the bet, acceptance criteria, non-goals, and evidence required to judge the outcome.

14 · DELIVERYShip it without losing the threadSequence the rollout with its gates, its approvals, and the signal that tells you the thing actually landed.

15 · PRODUCT DESIGNDesign from the job, not the libraryMake the surface and interaction calls that follow from your buyer and the job, instead of from whatever the component library offers.

16 · AI-NATIVE PRODUCTDecide where the model belongsPlace the model where it earns its keep, name what it must never decide alone, and set how you will judge it.

05 · GO TO MARKETStratOS picks the motion, plans the launch, and drafts the work. Then it tells you whether it worked.

17 · MOTIONPick the motion your buyer allowsChoose founder-led, product-led or sales-led from your actual buyer and sales cycle, not from what worked at a company you are not.

18 · LAUNCHLaunch with a reason to believeTurn the committed buyer, position, and proof into a coordinated first fortnight, not another disconnected launch plan.

19 · CONTENTTurn a position into work that shipsCreate posts, pages, and briefs from the message you committed to, with uncertain claims clearly marked before they go live.

20 · SALESWalk into the call already alignedGive your team the buyer context, objections, proof points, and qualification logic that the company has already agreed on.

Spend time building from expert knowledge.

StratOS gives your AI stack the hard-won judgment behind great product, growth, and go-to-market decisions, so it builds toward real customer value, not the average answer.

PLATFORM

From idea to execution, with strategy in the loop.

Customer calls, product data, revenue, feedback and market context go in. Clear positioning, product priorities, GTM decisions, experiments and build-ready specs come out. Every result returns as evidence, so each move is sharper than the last.

01 UNDERSTAND

See the company as it is.Start with your URL, deck, documents, calls, product signals, and revenue data. They become one connected picture of where you stand, not another folder of context.The first read establishes your stage, ICP, go-to-market motion, the assumptions that matter and what is still to prove. Every claim shows what it rests on: known, assumed, or to verify, and the knowledge autopropagates, building your knowledge graph.

SOURCES

ASK THE SYSTEM

Read my company.→ L1 · COMPANY READ

What is holding growth back?→ L2 · DIAGNOSIS + BETS

How do we price and win?→ L3 · POSITION + BETS

What do we build next?→ L4 · PLAN + MILESTONES

Ask the system…SEND ↵

L1READS & REPORTS

L2NAMES WHAT IS WRONG

L3TAKES A POSITION

L4SHIPS

CR-01Company readSUMMARY · GAPS

ICP-02ICPDIAGNOSIS · 2 BETS

MKT-03Market modelDIAGNOSIS · 1 BET

POS-04PositioningPOSITION · SET

PRC-05PricingPOSITION · 1 BET

SPEC-06SpecMILESTONES

GTM-07GTM planRISKS

PRI-08PriorityCUTLINE

⬧COMMITTED BET · GTMREAD BY 14 NOV2 OF 3 APPROVED

Mid-market lands faster than SMB

KILLFewer than 4 of 10 close

RECOMMENDED

Interview 5 stalled SMB dealsInterview 5 stalled SMB deals⬧MID-MKT✓

Pull close rate by segmentPull close rate by segment⬧MID-MKT✓

Draft the security one-pager⬧MID-MKTHELD

YOU APPROVE

APPROVEDAPPROVED

IN FLIGHT

LOGGED

Low-light detection PRD⬧SPECRUNNING

Pricing test · SMB tier⬧ICPLOGGED

02 SET THE WORK

Plan for your growthAs your company knowledge builds, the system proposes the next move: a decision to make, a bet to test, an action to take, or a deliverable to create.Challenge the recommendation. Add context. Ask what it changes, what it depends on, and what would prove it wrong.When you commit, it becomes a live decision. The system tracks the evidence, the deadline, the work it creates, and the result that confirms, changes, or ends the bet.

03 EXECUTE

Ship the workEvery action and deliverable leaves as an executable your tools can run.Each one carries the bet it serves, so the thing doing the work knows what the work is for, and the result knows which question it answers, and you stay in control.

⬧ BET · GTMMid-market lands faster

COMMITTED

ACTIONS · QUEUEDEVERY ROW CARRIES A BET

40-account target listACT-01 · HEY RICH

GTMAPPROVEAPPROVED

Open PR from specACT-02 · CLAUDE

GTMAPPROVE

Update pricing pageACT-03 · WEBFLOW

PRICEAPPROVE

DELIVERABLESEACH DECLARES ITS PARENT

GTM planGTM-07

GTMCOMMIT

Low-light detection PRDSPEC-06

GTMCOMMITCOMMITTED

Audit-log migrationINF-12

NO BETCOMMIT

APPROVE

COMMIT

SEND TO YOUR AI STACK FOR EXECUTION

AWAITING APPROVALEXECUTING

AWAITING APPROVAL

AWAITING APPROVALEXECUTING

RESULT RETURNS TO THE BET → 04

PIPELINE

USAGE

DELIVERY

FEEDBACK

DOCS

IN

EVIDENCE RETURNS

L1

READS & REPORTS

L2

NAMES WHAT IS WRONG

L3

TAKES A POSITION

L4

SHIPS

CR-01

Company read

SUMMARY · GAPS0 SIGNALS · 5 SOURCES3 RE-READS QUEUED

ICP-02

ICP

DIAGNOSIS · 2 BETS

MKT-03

Market model

DIAGNOSIS · 1 BET

POS-04

Positioning

POSITION · SET

PRC-05

Pricing

BET LIVE · 30D

PRC-05

Pricing

SETTLED · INVALIDATED

SPEC-06

Spec

CURRENT

SPEC-06

Spec

RE-READ

GTM-07

GTM plan

CURRENT

GTM-07

GTM plan

RE-READ

PRI-08

Priority

CURRENT

PRI-08

Priority

RE-READ

3 BLOCKS FLAGGED · NOTHING REWRITTEN

LOOP CLOSES → 01 READ

04 · COMPOUND

Let the market answer.Your company keeps telling you what changed. Attio, Amplitude, Linear, Dovetail, and Confluence or your direct customer conversations return the evidence against the threshold you set.You see whether the bet held, failed, or needs to change. No chasing updates. No rebuilding the story from memory.Just a clear OS that tells you what worked, what did not, and what to do next to reach your growth goals.

PRICING

One simple plan

Build from a company record that learns.

$49/ MONTH

Start freeFULL ACCESS, NO CARDWHAT YOU GET

Full access to every agent, unlimited
Evidence, assumptions, decisions, and bets in one record
Refreshes and stale downstream work flagged automatically
Connect your calls, CRM, revenue and analyticsROADMAP
Build-ready specs, plans, pages, and briefs for your AI stack
MCP access for Claude Code, Cursor, and CodexROADMAP
No credits, no seats, no token packs

EARLY ACCESS · ROLLING COHORTS

Know what to build.Build with conviction.

Do not let AI ship the wrong thing faster. Before work starts, decide what is worth doing.

Built for founders, builders, product managers, developers and operators Built forfounderswho need every decision to be clear, connected, testable, and ready to execute.

A new species of strategy tool. Built to work in the terminal beside your coding agents, StratOS sets a new standard for how decisions become product: your agent builds from what you committed, and tells you where your code disagrees before anything ships.

Spend time building from expert knowledge.

From idea to execution, with strategy in the loop.

One simple plan

The tools do the work.StratOS decides what the work is for.

Know what to build.
Build with conviction.

Built for founders, builders, product managers, developers and operators
who need every decision to be clear, connected, testable, and ready to execute.

The tools do the work.
StratOS decides what the work is for.