ai-venture-studio 0.68.0

AI venture studio: a full-lifecycle multi-agent system that finds opportunities, writes PRDs, builds, launches, and measures — with typed claims and human gates

AI Venture Studio (avs)

AI Venture Studio is an open-source multi-agent system that takes one plain-language requirements document and plans, builds, tests, reviews, and ships the product — with a human approving every irreversible step. It also runs the loop around the loop: it finds what to build from your real user signals, sizes it honestly, writes the PRD with kill criteria, measures whether it worked — and forces the kill decision when it didn't.

Try it in two minutes, no API key:

uvx --from ai-venture-studio avs replay --demo

That replays a real review of this repo's own code from its vendored audit trail, offline — including the run where the pipeline escalated to a human instead of guessing.

This README cannot overclaim. Every quantitative claim in it is machine-checked in CI against claims/platform.yaml — a number that was not measured, or a superlative about anyone else, fails the build (ADR-U29).

Pick your door

One spine, three editions — narrowing presets over the same unchanged pipeline, never forks, never fewer checks (edition_lint refuses anything that widens):

• Solo founder / OPC 一人公司 — built around your attention: one product bet, a 45-minute weekly review with the kill criteria first, safe publishing defaults. FDRs in English or Chinese.
• Engineer — fixture-gated extension points; nothing unfixtured registers, and you're invited to try to break that.
• Enterprise / traditional industry — ships the procurement pack and the pilot-to-production contract; init refuses without a named gate owner.

For founders (no technical background needed)

Three steps to the browser UI:

pip install ai-venture-studio      # 1. install — `avs` appears on your PATH
export ANTHROPIC_API_KEY=...       # 2. the key that powers the build
avs studio myteam --profile web    # 3. serve the Studio for a new workspace

Then open http://127.0.0.1:8433. The Studio is localhost-only: no account, no signup, nothing leaves your machine except the model API calls. Profiles: web | miniprogram | app. Returning later is just avs studio from inside the folder. (Wrong Python environment? uvx --from ai-venture-studio avs studio myteam --profile web sidesteps it.)

One real run, unedited — a live provider, no mock, nothing composited. Plain language in; a plan comes back for confirmation; modules build with live per-task progress; the report says in your own words what works and what doesn't; the last frame is the product that run actually produced. That report reads "partly built" because some modules failed — it names them, keeps the rest working, and offers a retry per module. This is what a real run looks like, which is the only kind of demo this repo will ship (single screen · Chinese UI).

English is the default (--lang en); --lang zh gives the original bilingual UI for 微信小程序 founders. The page header switches the view — Founder, Engineer, Enterprise — and deeper views only add read-only detail.

Your entire input is an FDR: six questions answered in your own words. The one that produced the run above:

## 1. Who is this for?
The two of us running a small studio. Right now we track work in chat
messages and keep losing track of what is actually finished.

## 2. What do users do with it?
1. Anyone adds a task with a title and who it is for.
2. Anyone marks a task done.
3. We look at the open tasks and the finished ones separately, newest first.

## 3. Must-have features
Add a task. Mark a task done. See both lists.

## 4. NOT needed for now
Logins, due dates, notifications, a mobile app.

## 6. What does success look like?
The two of us stop tracking work in chat messages.

• If your FDR is unclear, the system asks — it never guesses (at most 5 questions a non-technical person can answer, in your language).
• One FDR = one thing. The first FDR is the smallest usable product; every later feature is its own small FDR. Small builds are more accurate and fail more debuggably.
• You confirm intent in plain language before anything is built, and get a build report in your language after — including every automated approval the machine made on your behalf.
• You can see what it is doing while it does it. Each module narrates its own steps as they start — working out how to build it, writing the code (attempt 2 of 3), running your tests, checking that it actually starts up — in the terminal and in the Studio. No percentages and no ETA: the system does not know whether attempt 2 will be the last one, and a made-up number is worse than an honest step.
• A module that fails says why. Not "the build gate failed" but the test that failed, or that the model's answer was cut off and the task needs splitting. Every attempt is kept for inspection.
• Real persistence out of the box (local SQLite; Supabase and WeChat Cloud 微信云开发 are guided options with credentials in a vault that never enters prompts).

The same flow runs in the terminal: avs create → avs preview → avs add → avs ship (full founder walkthrough).

What happens when you press build?

Fourteen gated stages across two loops, and every generative stage runs the same template: one writer, deterministic tools first, independent charter voters (each fixture-gated at 8 cases, ≥87.5% to register), a fresh verify pass per finding, a leader synthesis, and a gate — human wherever judgment is the point. Nothing auto-merges, nothing deploys autonomously, nothing publishes, nothing spends.

The full design is public — fourteen documents, cross-referenced to shipped code in the implementation map (open items are named, never silent): autoproduct-design.

A real run (unedited)

Real support tickets in, an evidence-gated product decision out. Signals, verbatim from the tracker:

s1  "I started a build and stared at the terminal for 40 minutes with
     no idea whether it was progressing or stuck"
s6  "how much will a typical month of builds cost me? I'm scared to
     leave autopilot running"

avs opportunity → market → prd → evidence. Every artifact below is unedited pipeline output from one real-provider run:

1. P0 turns the signals into grounded candidates (Gate PL0) — every claim cites its ticket verbatim, each carries a falsifiable hypothesis and a named cheapest test ("ship a clickable mockup to the 3 reporters", not "build an MVP").
2. P1's own voters attack the market case — the Sizing seat caught an ungrounded 0.15 affected-fraction inference; the dedicated Disconfirmation seat argues the other side of the same evidence. The deterministic gate had already blocked an earlier draft outright: 75% of its claims were model_inference against the 30% market ceiling — reasoning dressed as research doesn't pass.
3. P2 writes a PRD with its own death spelled out (Gate PL2) — kill criteria authored before anyone is attached, sibling candidates listed as non-goals by name, and a Planning task auto-generated for the metric nobody had instrumented yet.
4. P4 reads the cohort and refuses to flatter it:

build_progress_view_rate: 0.240   n=250   CI [0.191, 0.297]   window complete
verdict H-1: insufficient_evidence — the interval brushes the 30% kill
threshold; the honest output is the n it would take to know, not a win.

What has been measured?

Numbers below link to their evidence; the CI-enforced ledger entry for each is in claims/platform.yaml.

• Review benchmark (avs bench): recall 100%, precision 67% on 13 labeled cases against bars of 40% and 50% (cases, method).
• Product benchmark (avs product-bench): full FDR→product runs scored by independent behavioral probes executed against the built product (WebGen-Bench pattern), reported unaveraged across synthetic and real case sets. Synthetic cases (2026-07-23, n=1, claude-opus-4-8 writer): build 100%, probe pass 83.3%, clean review 100%. Real cases, run 5 (2026-07-26, n=1, pre-fix baseline): build 33%, probe pass 0%, clean review 17% — published because a benchmark you can only pass is marketing; one you can fail in public is evidence (full run history).
• Perf-lane calibration: 5 of 5 seeded defects caught (catch rate 100%) at the 3x relative-detection factor, loopback low-parity environment, 2026-07-26 (manifest).
• 1560 hermetic tests (uv run pytest, no network, no keys); every PR in this repo was reviewed by avs itself, and five of those reviews caught real bugs.
• A hermetic suite is not enough, and this repo says so. Twelve real defects were found in one day of running the product against one real requirements document, with every test green — two of them at the model-SDK boundary, where a mock is authored by the same person holding the same wrong belief. avs smoke makes four real calls per configured provider and is step 0 of every release (runbook).

How does it compare?

Orchestration SDKs and app builders are complements at a different layer — you can keep their mental models and still adopt this repo's lifecycle, gates, and evidence discipline. No scores, no adjectives; just the layer split:

	Layer	Opinion held	Opinion delegated
Orchestration SDKs (LangGraph, CrewAI, agent SDKs)	build agents	how agents compose and communicate	what your team's lifecycle, gates, and evidence standards are
Chat-to-app builders	generate an app from a conversation	speed from prompt to running code	what happens after: review depth, launch honesty, the kill decision
AI Venture Studio	run a product lifecycle	which decisions need a human, which claims need evidence, and when a product must die	which model families do the judging (swappable seats)

Limitations

• The outer loop runs end-to-end and survived its first real-provider smoke, but its release bar is honest: it is unproven until a real Gate PL5 records a real kill or pivot on a live cycle.
• Cloud services are guided, not auto-provisioned; deploys generate artifacts + instructions, and the button stays yours until you arm a policy that says otherwise (ADR-031 — disarmed by default, attributed, expiring).
• 小程序 runtime verification (avs mp-runtime) needs the WeChat DevTools desktop app plus a one-time human toggle (Settings → Security → Service Port), so it runs on macOS/Windows and never in CI; every missing precondition is a visible skip naming its remedy, never a silent pass. Page-level unit testing still needs miniprogram-simulate; pure-logic modules are gated via node --test.
• Single-machine operation; crash recovery resumes reviews, deploy reviews, and incidents from their checkpoints, but multi-instance supervision remains the documented upgrade path.
• Multi-tenant isolation is filesystem-and-routing level, not OS-level (ADR-030); if that is in your threat model, run one process per tenant.

For developers

pip install ai-venture-studio   # the command is `avs` (`autoproduct` kept as an alias)

The repo ships no keys, no proxy, and no metered backend: every provider call bills the keys in your environment, and every provider errors loudly if its key is missing rather than running half-armed. OPENAI_API_KEY is optional but recommended — it puts a different model family in the security voter seat, breaking same-family self-preference when Claude reviews Claude-written code. Setup, env vars, and operations: RUNBOOK.md.

The CLI at a glance (every stage is one command; full table in the RUNBOOK)

opportunity · market · prd · evidence (+ *-approve)	the outer loop as one-command stages; human decisions recorded at gates PL1/PL2
discover / plan / spec / build (+ *-approve)	inner-loop upstream stages, gates U1–U4; scr is the only legal way to change a built spec
review · resume · replay · recover	the review pipeline, HITL, audit trail, crash recovery
deploy-review · triage [--fix]	deploy gate and production maintenance
bench · product-bench · voter-gate · compound --pr	the benchmarks, voter registration gates, and the weekly compounding loop
automerge · deploy-execute	exist but stay disarmed until a human writes an attributed, expiring policy (ADR-031)
readiness · attest · cab-package · sweep	the enterprise adoption surface: substrate ladder, attestation ledger, change control, the janitor
mp-runtime	opens a built 小程序 in WeChat DevTools, visits every registered page and screenshots it — the blank-page check the static gate cannot make (pipeline guide)

FAQ

Can it build an app from a plain-English description?

Yes — that is the founder flow: one FDR (six questions in your own words, English or Chinese) goes in; a planned, built, tested, and reviewed product comes out, with your plain-language confirmation gate before any build starts. The product benchmark exercises exactly this path end to end.

Do I need to know how to code?

No. The Studio UI asks for your product description in your own words, asks clarifying questions when it is unclear, and reports back in plain language — including an acceptance walkthrough you can click through. You never have to read the code it writes (it's yours, though).

What does it refuse to do autonomously?

Merge to main, deploy to production, publish, send, spend money, fabricate user evidence, or close a fired kill criterion. Each of these requires a recorded human decision; merge/deploy automation exists but is disarmed until a human writes an attributed, expiring policy naming exact branches.

Does my data leave my machine?

Only model API calls, billed to your own keys. There is no account, no proxy, and no metered backend; telemetry is off by default and aggregate-only if you opt in (avs telemetry show prints the exact payload before anything sends). Credentials live in a vault layer that never enters prompts.

What happens when a build fails?

Passing its own tests does not get a module past the build gate: a web product must actually start and listen on its port, and a 小程序 must be openable in WeChat DevTools — app.json present, every page registered, and every relative require chain resolving to a file inside the mini-program root, because a module that throws at require time is a page that renders blank. Both checks exist because real runs produced green suites over products that could not load.

For 小程序 there is a rung past that, on your own machine: avs mp-runtime opens the built product in WeChat DevTools, visits every registered page and screenshots it. A page counts as broken when its screenshot is a single flat colour — because "the page opened without throwing" is not evidence: a page whose JS died before Page() still opens, still sits on the page stack, and still renders pure white. That check exists because a build was reported as seven pages rendered while three of them were blank. The pipeline guide has the four rungs and the failure that justified each.

The run retries its own mechanical failures first: one bounded pass, in dependency order, with the previous attempt's diagnosis handed to the writer — so a retry is a different attempt, not a replay. What still fails is preserved for post-mortem, the rest of the product keeps working, and the Studio offers per-module retry — an interrupted build resumes from what is already built instead of re-paying it. Review verdicts and gate records stay on disk as YAML you can replay.

Can it run up a surprise bill?

Signal s6, verbatim: "how much will a typical month of builds cost me? I'm scared to leave autopilot running." The answer is visibility plus your provider's own controls. Every call is billed to your key or subscription — the framework holds nobody's keys and never spends money on your behalf — so spending limits belong where the billing actually happens: your provider account, whose limits see all usage on the key and cannot be bypassed by anything here. What the framework owes you is the number, and it delivers it everywhere money is decided: every build report ends with what the run cost, as arithmetic; avs cost prints the month per model; the Studio shows spend on the confirm page, before the first dollar. Prices are published list prices with a source and a date (avs prices), ranges resolved upward so the estimate is a ceiling, and a model with no sourced price keeps the total honestly labelled a floor — never counted as zero. There is deliberately no framework-side spending cap (ADR-032): it would duplicate your provider's control and mislead subscription users whose tokens don't map to marginal dollars.

How is this different from an orchestration SDK like LangGraph or CrewAI?

Different layer: SDKs give you primitives to build agents; this is an opinionated product lifecycle that happens to be run by agents — with the opinions (gates, evidence rules, kill criteria) enforced by deterministic code, and the judging seats swappable across model families. See How does it compare? above.

Status

Current release: see the PyPI badge and CHANGELOG.md; the release-by-release record back to v0.8 is in docs/release-history.md. Next milestone: the v3.0.0 design gate — the launch PRD's kill criterion needs four consecutive logged attention weeks, and a human records the decision (runbook).

---

MIT · design docs: autoproduct-design · operations: RUNBOOK.md · security: SECURITY.md