Find the open-source solution that fits.

venture-compass is a Claude Code co-pilot for solo AI-built micro-ventures. It keeps state in VENTURE.md, watches kill dates, live metrics, and landscape freshness, and convenes independent adversarial lenses when a gate is due. The stages cover whether to take on an idea, success and kill criteria, checkpoint verdicts, and a constrained pivot, returning GO, PIVOT, KILL, or SCALE with a default bias toward killing on time. Project-local metric collectors run only after their source is explicitly marked verified.

Ethical Response Reviewer checks a draft AI response before delivery and chooses approve, revise, refuse, clarify, or escalate. It explains the decision, supplies a safer rewrite or escalation note when required, and emits machine-readable audit tags for later logging. The skill is aimed at high-risk contexts such as law, medicine, finance, employment, public services, minors, and sensitive data. It is a structured risk layer, not a guarantee or substitute for professional advice, human responsibility, access controls, or platform safety.

shopping-aggregator triages a purchase by product, region, budget, urgency, and condition, then delegates live research across thirteen consumer-shopping domains. It ranks landed cost rather than sticker price, pins the exact variant and timestamp, checks stock and coupon claims in the cart, and surfaces seller trust tiers, coverage gaps, and disagreements between snapshots. It is an orchestration layer rather than a price-fetching engine, so missing research tools become explicit gaps. The skill recommends what to buy but never completes the purchase.

agentic-commerce-news scans the previous seven days of public X posts, industry media, and VC or platform announcements for agentic-commerce startups, launches, funding, partnerships, and influential opinions. It filters for credible endorsement, classifies items across a nine-layer commerce stack, and writes sourced cards, an overview table, and weekly takeaways. Quality rules reject stale, duplicated, or purely sponsored items and require dates from source pages. It needs a WebSearch-capable agent; persistent scheduling is optional and environment-specific.

spark-forge packages local ComfyUI generation on NVIDIA DGX Spark as Claude Code plugins for images, 3D models, music, video, and gated publishing. FLUX.2, Hunyuan3D, ACE-Step, WAN, and HunyuanVideo are selected through model profiles and quantization options, while setup skills provision the stack and private S3 or CloudFront delivery. It requires a ComfyUI container, Docker, AWS CLI for publishing, and downloaded model weights. Model licenses vary and some are non-commercial, so the repository’s terms must be checked before use.

product-prover reviews PRDs, feature specs, architecture documents, and design proposals before coding. It extracts states, transitions, actors, boundaries, invariants, safety, liveness, and rollback conditions, then ties each finding to a source location, trigger, consequence, and concrete fix. Unsupported conclusions stay marked as missing specification rather than invented. Full-spec and focused cross-link passes are available, with findings persisted by status; this repository is a read-only mirror of a larger pack.

cross-stake reads CROSS Chain Mainnet network state, validators, wallet delegations, rewards, and unbonding history, then can submit delegate, undelegate, claim, and compound actions behind explicit confirmation gates. Read-only checks do not require a private key; write settings can enforce gas reserves, notional limits, and confirmation thresholds. Managed signing through Vault, KMS, or HSM is supported as an alternative. The skill targets CROSS Mainnet and should be paired with treasury and risk controls before real writes.

logo-maker makes an agent study the project, audience, and purpose before distilling one core idea into a simple, meaningful mark. It generates many candidates, checks each in light and dark modes plus a 16px row, then recommends only the strongest few and exports SVG, favicon, PWA, or related identity assets. Its rules favor flat two-color construction with one meaningful accent and reject generic system-font letters, rainbow gradients, clip-art objects, and app-tile widgets.

social-fact-check extracts factual claims from a draft social post, separates them from opinion, and verifies each claim against a strict source hierarchy led by primary records. It checks that cited studies, laws, dates, document numbers, statistics, and quotations exist and support the wording, then returns a verdict table, required changes, a corrected version in the brand voice, and residual risks. It is built for teams that cannot afford public errors and relies on web search for current evidence.

survey-linter checks survey questions for leading wording, double-barreled items, presuppositions, absolutes, vague quantities, recall burden, unbalanced scales, missing options, priming, and fatigue. Its rule engine is deterministic and points each finding to exact wording, with a score whose arithmetic can be reproduced without trusting a hidden model. English and Simplified Chinese are supported, along with text, Markdown, CSV, and common form JSON formats. It runs locally, offline, in a browser, CI, or as a Claude skill.

weekly-news-seo gathers a niche’s weekly stories, curates a small shortlist, and drafts original articles for classic search and AI answer surfaces. Each article adds context beyond its sources and can include direct-answer blocks, question headings, sources, structured data, and internal links. You approve or reject every draft before it is written into the site, committed, and deployed through Vercel; the workflow is intentionally low-volume rather than an auto-publisher.

Dashmotion converts a plain-language workflow or a supported Mermaid diagram into one self-contained animated HTML/SVG file. Flow mode shows execution moving through stages and branches, while architecture mode adds component boundaries and animated request journeys. The output is small, dependency-free, editable through follow-up natural-language requests, and accessible with pause controls and reduced-motion handling. Sequence, class, ER, and gantt diagrams are outside the supported input set.

report-kit helps an author produce a single HTML report that recipients can open in Edge or Chrome without installing anything. The page supports in-place editing, saving changes through the browser File System Access API, and PNG export through an embedded html2canvas copy, so it works without a CDN. The author needs Claude Code and Python for bundling; Firefox can open the file but falls back to downloads for save and image actions.

voz treats writing as craft rather than as a detector or score. It removes clusters of machine-like vocabulary, symmetry, inflated phrasing, and formatting habits, then applies one of nine audience registers such as technical, journalist, legal, teacher, or marketing. You can invoke it per task or wire the default voice into Claude Code through a project instruction and SessionStart hook. Its word list is dated and a single tell is never treated as proof.

professional-repo-docs bootstraps a GitHub repository’s documentation from a natural-language request, including README variants, a suitable license, community-health files, issue and pull-request templates, AGENTS.md, and llms.txt. It detects the project stack to fill real build and test commands, copies canonical license and conduct text, and creates only missing files. The Node-based engine never auto-commits; the optional GitHub CLI is only needed to enable Discussions.

blockpad-documents helps Claude Code create or edit `.bpad` files containing units-aware equations, tables, two-dimensional drawings, and solver sections that recompute in Blockpad. It can also convert a Word file, PDF, Markdown document, or photographed hand calculation into a live calculation sheet. A Python validator checks XML structure, schema details, and known formatting or formula traps; the integration targets Blockpad 1.8.x and is unofficial, so the desktop app remains the final visual check.

thai-pdf-booklet renders Markdown into individual Thai PDFs or a bound booklet with a full-bleed cover, back cover, table of contents, page numbers, and per-page credit text. Its portable Node CLI bundles Chromium and embeds Noto Sans Thai, so macOS, Linux, and Windows users do not need pandoc, poppler, system Chrome, or system fonts. A native variant can use an existing Chrome, pandoc, and poppler installation instead of downloading Chromium.

Google Ads Custom Affinity Builder turns a product, market, persona, and user-supplied sub-themes into a multi-tab XLSX with keyword, website, video, channel, and app signals. It uses yt-dlp for view-aware video search, enforces a 10K-view floor, checks a channel’s recent uploads rather than trusting one viral video, and sends candidates to an independent checker. The workbook is designed for Google Ads entry, with tab names constrained by Excel’s 31-character limit.