ideaudit

studio.inite由 inite.studio 發布官方發布者v1.1.0更新於 Oct 24, 2026

The scoring behind an audit allowed to say no. Twenty deterministic tools, offline, no account.

已驗證Streamable HTTP可網頁執行資料與分析商業與電商

概覽

AI 產生的概覽

面向創業點子稽核的決定性評分工具：21 個純數學函式，把市場、資金與訊號輸入換算成分數與 GO/KILL 結論。

可以做什麼
  • 一次呼叫了解伺服器用途以及帳號能帶來什麼
  • 計算結論：dealbreakers v2、交叉原型矩陣、推導出的終止條件
  • 組合評分：LRS 綜合分（v1 與 v2）、急迫度綜合分、12 項集合評分
  • 評估市場形態：進入門檻、多來源 TAM 共識、搜尋熱度（v1 與 v2）、社群痛點
  • 評估資金面：單位經濟驗證、變現能力、預算證明、PPC 投放訊號
  • 評估訊號：募資動能、招聘需求、X/Twitter 訊號、建置複雜度
適用情境
適合希望用透明、可重現的算術而非模型判斷來為創業或產品點子評分，或想用自己的資料核對稽核數值的情境。若只想驗證演算法本身，僅安裝本機套件即可。
執行需求
可透過 npx 從 npm 套件 @inite/ideaudit-tools 以 stdio 在本機執行，也可使用 api.inite.studio 的託管 HTTP 端點。本機伺服器無需帳號、金鑰、網路或模型呼叫。託管伺服器需在標頭 Authorization 提供來自服務商設定頁的 Bearer 權杖；未帶此標頭時端點以訪客通道提供決定性評分工具。
安裝前請注意
本機伺服器不儲存任何內容：傳入 auditId 會被拒絕，結論只存在於對話中，之後無法重新開啟。持久化、第三方市場資料與知識圖譜需要託管伺服器與帳號。託管端點接受 Authorization Bearer 權杖，屬於需妥善保管的憑證。評分品質取決於你提供的輸入。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 ideaudit，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/studio.inite%2Fideaudit-tools/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "ideaudit": {
      "type": "http",
      "url": "https://api.inite.studio/mcp"
    }
  }
}

README

ideaudit-tools

The deterministic scoring behind ideaudit, as a local MCP server. No account, no key, no network, no model calls — just the arithmetic.

bash
claude mcp add ideaudit-tools -- npx -y @inite/ideaudit-tools

Listed in the MCP registry as studio.inite/ideaudit-tools, which carries both ways in: this package over stdio, and the hosted server over HTTP. Installing straight from this repository — npx -y github:inite-ai/ideaudit-mcp — also works and runs the same code.

Why this is published

The product's claim is that it is an audit allowed to say no. A verdict nobody can inspect is a verdict nobody has to believe, so the part that produces the number is open — including the number that says no. You can read how three lens scores become KILL, disagree with the thresholds, and argue about the arithmetic rather than about our word.

It is also the honest shape of the system. The server makes zero model calls; judgement lives in the skills, determinism lives in the tools. These are the tools.

What runs locally

Twenty tools, every one a pure function of its input:

The verdictcompute_dealbreakers_v2, compute_crossed_matrix, derive_kill_criteria
Composite scoringcompute_lrs_composite, compute_lrs_composite_v2, compute_urgency_composite, compute_collection_scores
Market shapecompute_barrier, compute_multi_source_tam, compute_search_velocity, compute_search_velocity_v2, compute_social_pain
Moneyvalidate_unit_economics, compute_monetization, compute_budget_proof, compute_ppc_spend_signal
Signalscompute_funding_momentum, compute_hiring_demand, compute_x_signal, compute_build_complexity

compute_dealbreakers_v2 is the one that decides. Stage-aware weights, confidence-weighted lens scores, and a deliberately risk-asymmetric verdict: GO needs a score ≥ 80 and zero red flags and average confidence ≥ 0.6, while the KILL gate is observer-invariant — a founder profile shifts every threshold above the gate, and none of them below it. You cannot buy a better answer by describing yourself more favourably. There is a test that says so.

Same input, same output as the hosted server — because it is the same specification object, imported rather than reimplemented. A test asserts that identity rather than comparing two copies, so the two cannot drift.

Where to look

Be told where the numbers are rather than hunt for them:

  • src/specs/*.ts — one file per tool: what it accepts, what it means, what it returns. compute-dealbreakers-v2.ts is the verdict.
  • src/server.ts — the whole server, including what it refuses.
  • src/__tests__/parity.test.ts — the claims above, as assertions. The observer-invariant KILL gate is checked against all five archetypes.
  • dist/bundle.js — the arithmetic, un-minified. computeBarrier is a dozen lines and every constant is visible: 0.6 off the score per direct competitor, 0.2 per adjacent, six points of range spent on SERP noise, and labels at 18 / 12 / 6. Disagree with those numbers, not with our word.

That last one is a build artefact and reads like one — the commented source of the maths lives in the monorepo this repository is mirrored from, and is not here. If you want the reasoning behind a threshold rather than the threshold itself, open an issue and ask; that is a cheaper conversation than either of us pretending.

What needs the hosted server

Not withheld to make a point — these genuinely need somewhere to keep things and someone to pay for them:

  • Persistence. Saving an audit, finalising it, reading it back, comparing two of them, the dashboard. A verdict that exists only in a chat transcript is one nobody can re-open or defend later.
  • Data. SERP, keyword volume, funding rounds, hiring signals, social mentions. Those are metered third-party APIs.
  • The knowledge graph. Entities, evidence, contradictions across audits.

That boundary is enforced rather than documented. The hosted compute_dealbreakers_v2 takes an auditId; this one refuses it, by name, instead of letting zod strip it silently and handing you a correct verdict plus the false belief that it had been written down.

$ compute_dealbreakers_v2 {"auditId": "…", …}{ "error": "unknown_arguments", "unknown": ["auditId"],  "message": "This server computes; it stores nothing. …" }

The hosted server

bash
claude mcp add ideaudit --transport http https://api.inite.studio/mcp \  --scope user --header "Authorization: Bearer $IDEAUDIT_API_KEY"

It serves these twenty tools and everything above them. Sign up at inite.studio; --scope user matters, or the registration is bound to one directory.

Using both

They compose. Register the local server for the arithmetic and the hosted one for everything else; tool names do not collide because the local set is a strict subset served under the same names, so whichever you register wins for those twenty and the rest resolve to the platform.

If you only want to check our maths against your own numbers, the local server alone is enough and always will be.

Contributing, and a warning about it

This repository is a projection. It is rebuilt and force-pushed from the monorepo that develops it every time the package changes, which means a pull request opened here would be overwritten by the next sync — through no fault of yours, and with no notification. Saying so is cheaper than letting someone find out by losing an afternoon.

Issues are the right channel and they are read: a threshold you think is wrong, a number you cannot reproduce, a tool that accepts something it should refuse. Bring the input you used — every tool here is deterministic, so a disagreement about an answer is always settleable.

Contribute with an AI agent

[Open in GitHub Codespaces]

You don't need to be an engineer. Open the repo in your coding agent and say /contribute (Claude Code) — or "read AGENTS.md and help me report …" in any other agent. It runs the tool on your input, finds the constant involved, and files an issue a maintainer can settle — an issue, not a pull request, for the reason above. AGENTS.md also explains why dist/ is committed and how node scripts/verify-dist.mjs checks it against src/.

Licence

Apache-2.0. See LICENSE.

來源：README.md，提交 822ac38

工具

21

get_started

What this server is, what it will do for you right now without an account, and what an account adds. Call this first if you have no API key — it answers in one round trip instead of sending you to a website.

無需輸入參數。

compute_barrier

Compute barrier_score (0-24) + label (PRISTINE/OPEN/COMPETITIVE/CROWDED) from competitor counts + SERP noise fraction.

輸入參數serpNoisedirectCompetitorCount必填adjacentCompetitorCount

compute_budget_proof

Compute budget_proof_score (0-10) + label (STRONG/CONFIRMED/WEAK/ABSENT) + purchase_intent_pct from pricing hits + review-site hits + intent mentions.

輸入參數hasNamedPricingpricingHitsCount必填reviewSiteHitsCountpurchaseIntentMentions

compute_build_complexity

Compute build_complexity_penalty (0-10, higher = worse) + per-factor breakdown. Hard tags: ml/realtime/blockchain/hardware/compliance/custom-ai/regulated/on-device-ai/iot.

輸入參數externalApisCount必填integrationsCountstackComplexityTags

compute_collection_scores

Compute 12 deterministic collection scores (0-100) + badges + death reason for an enriched idea. Pure math. No external calls.

輸入參數analysisId必填enrichedData必填

compute_crossed_matrix

Crossed-product audit explorer. Same input as compute_dealbreakers_v2 — returns substrate verdict (no-observer baseline) + crossed verdict (when observer supplied) + a 5-row matrix of {solo, cofounded_technical, cofounded_business, domain_expert, serial} archetype verdicts. Never persists; meant for the dashboard "view as [archetype]" dropdown and for previewing a verdict before committing to it.

輸入參數stage必填sectorobserverlensScores必填stageProbabilitieshasMajorContradictionunresolvedContradictions

compute_dealbreakers_v2

Methodology v2 dealbreakers — stage-aware weights + confidence-weighted lens scoring + risk-asymmetric verdict (GO requires score≥80 AND zero red flags AND avg confidence≥0.6). Optional `observer` triggers the crossed-product pipeline: substrate verdict (no-observer baseline) PLUS crossed verdict (observer-perturbed weights, risk-tolerance shifted thresholds) PLUS 5-row archetype matrix. The KILL gate (≥2 blockers / score<50) is observer-invariant — fatal stays fatal.

輸入參數stage必填sectorobserverlensScores必填stageProbabilitieshasMajorContradictionunresolvedContradictions

compute_funding_momentum

Compute funding_momentum_score (0-10) + badge (HOT/WARM/COOL/COLD) from tier-weighted funding-article hit counts.

輸入參數hitsByTier必填recent30dHits

compute_hiring_demand

Compute hiring_demand_score (0-10) from priority-weighted ATS site hit counts (use registries/hiring-sources for priorities).

輸入參數sites必填

compute_lrs_composite_v2

LRS composite v2 — 6 components (SV, Pain, Barrier, Monet, X-Signal, Budget-Proof). Default Python weights 0.18/0.22/0.18/0.14/0.18/0.10 sum=1.0. Returns BOTH weighted score and equal-weight baseline (per OECD Handbook + Greco 2018 — equal-weight is defensible default when no outcome calibration exists). buildComplexityPenalty 0-10 subtracted from score. sectorProfile (ai_native/creator/crypto) opt-in reshuffles SV→0.16, X→0.20. Labels: THE_ROAR (≥80) / PROMISING (≥60) / EXPERIMENTAL (≥40) / WEAK_SIGNAL (<40).

輸入參數barrierScore必填xSignalScore必填sectorProfilesocialPainScore必填budgetProofScore必填monetizationScore必填searchVelocityScore必填buildComplexityPenalty

compute_lrs_composite

Compose lrs_final_100 (0-100) + label (WEAK/EMERGING/GOOD/STRONG/ELITE) + leaderboard_eligible flag + sub-percent breakdown. Weights: sv 0.25, sp 0.30, barrier 0.25, monetization 0.20.

輸入參數barrierScore必填socialPainScore必填monetizationScore必填searchVelocityScore必填

compute_monetization

Compute monetization_score (0-21) + label + has_pricing_anchors from pricing anchors + model tags + deal cycle hint.

輸入參數dealCyclemodelTagspricingAnchorsCount必填

compute_multi_source_tam

Multi-source TAM consensus. Pass 2-3 sources of market-size text. Optional `estimateYear` per source — when supplied, the result includes yearRange and a hasStaleData flag (true if the span exceeds 5 years). Outliers are dropped by modified Z-score over the median absolute deviation when n≥4. Returns the extracted dollar amounts + consensus median + an agreement score 0..1, where 1 means every source lands within 20% of the median.

輸入參數inputs必填

compute_ppc_spend_signal

Wave 5 N.4 — compute ppc_spend_score (0-10) + label (STRONG/CONFIRMED/WEAK/ABSENT) + market_saturation from PPC traffic projection (avgCpcUsd, totalMonthlySpendUsd, optional competitorBidders + competition). Feed numbers from dataforseo_ad_traffic.

輸入參數avgCpcUsd必填competitioncompetitorBidderstotalMonthlySpendUsd必填

compute_search_velocity_v2

Search velocity (0-25) v2 — canonical 0.40*volume + 0.30*trend + 0.20*intent + 0.10*geo. CRITICAL: externalVolumeNorm MUST come from external sources (Amazon BSR / app store installs / job-board postings) — NOT the Trends timeline (would double-count, since Trends is itself normalized 0-100 within window). trendNorm is derived internally from trendsTimelineValues. Trends peak<50 zeroes the trend component (Yotpo SEO floor). Optional daysSinceLastSignal applies exponential freshness decay (search half-life 90d).

輸入參數intentNorm必填geoSpreadNorm必填externalVolumeNorm必填daysSinceLastSignaltrendsTimelineValues必填

compute_search_velocity

Compute search_velocity_score (0-25) from Trends timeline values + rising queries count + geo region count.

輸入參數geoRegionCounttimelineValues必填risingQueriesCount

compute_social_pain

Compute social_pain_score (0-30) + total mentions + dominant perspective (business/consumer/trend/mixed).

輸入參數painMentions必填categoryCountsintentMentionsurgencyMentions

compute_urgency_composite

Compose composite_urgency_score (0-10) + badge (LOW/MEDIUM/HIGH/VERY_HIGH/EXTREME) from 3 sub-scores: news, pain, hiring.

輸入參數newsSignalScore必填painSignalScore必填hiringSignalScore必填

compute_x_signal

Compute x_signal_score (0-20) + recency share + positivity rate from X/Twitter mention counts.

輸入參數mentionsCount必填recent7dCountfounderMentionssentimentNegativesentimentPositive

derive_kill_criteria

Derive a falsifiable, data-driven list of kill criteria from upstream signals — the outputs of validate_unit_economics and compute_dealbreakers_v2, plus an ICP drift count. Returns one row per rule with {rule, threshold, status, evidence?}, where status is tripped_now / monitor / cleared. Replaces prose kill criteria, which are tautologies that can never fire.

輸入參數unitEcondealbreakersicpDriftCount

validate_unit_economics

Sanity-check a unit-economics row before publishing it in a business-model slide. Catches the math-drift class of failures (customers × ARPU ≠ revenue), enforces the LTV/CAC ≥ 1.5 floor, the cohort-positivity check, and CAC payback bounds. Returns {ok, errors[{rule, severity, detail}], derived{ratios}}. Skills MUST regenerate the row when ok=false (block-severity errors); warn-severity errors should be surfaced in the final report but do not gate publication. No LLM calls.

輸入參數cacltvarpu必填customers必填grossMarginmonthlyChurnannualRevenue必填

版本歷史

1
  1. v1.1.0最新Sep 16, 2026

評分

已滿足適用檢查項的 90%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分
  • 已宣告認證方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

    請在儲存庫或註冊中心條目中宣告授權條款。