Find the open-source solution that fits.

claude-credit-audit scans CI workflows and scripts for headless Claude calls, Claude Code Actions, and Agent SDK usage, then estimates monthly cost, per-call ranges, and the day a plan could run out of credit. It parses workflow triggers, model identifiers, and plan limits, and can calibrate token tiers from local session history. The project notes that the referenced billing change is currently paused, so the forecast describes a possible future impact rather than a live charge.

Use this Claude Code skill after the script or brief is settled to build a director treatment for commercials, corporate films, or brand videos. It structures the presentation around visual, audio, and emotional touchpoints, then can generate storyboard imagery, references, music options, and assembled Canva and PPTX files. Image, Suno, and Canva integrations are optional; missing tools fall back to prompts and placeholders rather than invented assets.

pg-extension-lab gives PostgreSQL extension authors an isolated harness for development, regression testing, benchmarking, and performance experiments. Its references cover C/PGXS and Rust/pgrx extensions, accelerator paths, microservice boundaries, matched-recall benchmarks, and Pareto-style trade-offs so a fast number is not mistaken for a fair result. Reusable templates keep hypotheses and reports tied to raw artifacts, and the package installs as a Claude Code plugin.

Stock Industry Chain takes a company, ticker, or theme and organizes upstream and downstream companies for Chinese, Hong Kong, or US markets into a shareable HTML map. Clickable nodes open company cards and TradingView charts, while a second view can sort individual stocks by market measures. Claude performs the research and the page builder inlines the resulting JSON, so company relationships, tickers, and market data should be checked independently; the project is not investment advice.

codex-collab gives Claude and Codex one append-only Markdown journal so neither agent needs to relay the other’s context by hand. Both read the index and the current day at startup, then record what they did, verified, touched, and leave open after each round. Claude is the default lead for planning and implementation while Codex supplies an independent review, but both are peers in the journal. Delegation requires a Codex MCP server.

ecom-klaviyo-email takes one product or collection link and turns it into an editable Klaviyo campaign draft, using brand context, account performance, customer language, competitor research, and a product-referenced hero image. It returns subject-line variants, a preheader, audience prompts, send-time guidance, and local working files, while inserting the store’s reusable header and footer. The audience remains a placeholder and the skill never sends or schedules a campaign; provider keys should be stored outside plaintext dotfiles.

cooking-master turns a dish name, a list of ingredients, or an imported recipe into a tailored cooking plan. It asks about people, equipment, time, taste, restrictions, and pantry contents, then returns shopping guidance, timed steps, cues for judging progress, and common failure fixes. It can coordinate several dishes around available burners and remember a kitchen profile. The skill depends on Claude and may search online when its recipe library has no match.

lookdev creates a live browser studio where a person tunes visual output with sliders, color pickers, drag handles, crop controls, and inline annotations while the render updates immediately. The chosen state can be written back as reproducible JSON, settings, code, or 3D metadata, so the deliverable is more than a screenshot. It covers image, color, typography, layout, animation, and material work, and also supports prose and media comments.

deterministic-design renders the actual interface before judging it. A spatial audit measures the ink-weighted centroid against the optical center, checks the declared grid, spacing, alignment, contrast, and touch-target size, and draws findings back onto the screenshot. A separate fresh-eyes vision pass evaluates usability against Nielsen’s heuristics and produces a prioritized fix list. The skill composes with other design tools and adds measurement rather than replacing visual taste.

Render several clearly labeled variants into one contact sheet or motion sequence, then ask a vision or video model to score them and suggest the next parameter values. The Claude Code skill structures a coarse-to-fine loop with calibration anchors and an early-exit rule, reducing manual trial and error. Its judgment is still an opinion, so personal taste and brand choices should remain human decisions.

unigui-framework-skill distills 116 official uniGUI 1.60.0 documents into mental models, decision heuristics, VCL-to-uniGUI mappings, anti-patterns, and a deployment decision tree. It helps Delphi developers reason about the shift from desktop-style code to stateful web applications, and supports architecture design, code review, migration, performance work, and Standalone, Service, ISAPI, or HyperServer choices. The material is Delphi-focused, may lag newer releases, and has limited coverage of mobile controls and advanced APIs.

Not So Fast checks whether an agentic decision was verified with enough independence for its difficulty and material impact. A self-only critic cannot approve a high-materiality, hard-correctness call; the guard instead requires an independent check or escalates when none is available. It ships as a Claude Code plugin, a SKILL.md folder, and a dependency-free Python guard with deterministic verdicts. The project presents this as a reference contract, not an empirically validated performance mechanism.

validate-idea gathers search, community, and competitor signals from Chinese and English markets, then has Claude score the results and produce a bilingual decision report. Seven fetchers fail independently, and data-poor dimensions are kept low rather than filled with invented evidence, so the comparison can expose when demand is already served in another market. Firecrawl, Zhihu, Brave, or proxy configuration is optional by source. The project supplies the data layer and uses the running Claude for scoring, so it needs no separate LLM API key.

Dowsing turns an item, time, and loss context into a structured search report with a primary direction, prioritized locations, movement clues, a findability tendency, and concrete next steps. Its interpretations change across home, public, transit, and lost-pet scenarios rather than assuming every loss happened indoors. The project frames these results as a search heuristic for checking overlooked places, not as a guarantee or substitute for practical action.

html-deck treats HTML as the editable source for fixed 1280×720 slides, renders each section in a local Chromium-based browser, and assembles the screenshots into a 16:9 PPTX or a folder of PNGs. It suits dense, consistent presentations with code, charts, tables, formulas, or CJK text. The important trade-off is that the PowerPoint contains one full-slide bitmap per page, so text is not selectable there; Node 18 and a Chromium browser are required.

auditoria-ley-21719 examines a project’s stack, database models, migrations, and forms to identify personal or sensitive data, then checks 17 categories under Chile’s Law 21.719. It produces a prioritized Markdown report with file-and-line references and labels each item as implemented, partial, missing, not applicable, or requiring confirmation when the repository cannot show the answer. The package is a technical compliance aid, not legal advice, especially for high-risk processing.

magdil-rosh prevents an agent from shipping only the happy path. Before building, it loads the checklist relevant to the request, asks up to three questions at the forks that change the design, and reports what is covered, deferred, or still open. The WalterCare edition adds project-specific checks for its environments, data layer, compliance, and locales; trivial lookups and one-line edits are intentionally skipped.

Forge’s magdil-rosh routes a feature or non-code deliverable to a focused checklist, loading only the material needed for that request. It asks up to three scope questions, covers each relevant row or names the deferral, and reports tersely so the agent does not stop at the happy path. Project-specific sections can be adapted to the project’s backends, compliance requirements, locales, and design system; no API key or extra configuration is required.