Browse agent skills

Can a weaker AI catch a stronger one hiding a secret objective? Alignment audits in the terminal, in Claude Code and in CI, with semantic-entropy scans and an oversight-gap study on the Claude 5.5 family.

Ollama provider for Firekylin AI services

Config-driven typed classification for Pi via Microsoft Decision-1, OpenAI Decisions, Perplexity, TypeSafe Jev and Cloudflare Clef, with selective fallback and explicit comparisons.

Fit tool outputs into token budgets while preserving useful structure.

Google Gemini provider for Firekylin AI services

README generator skill for Claude Code. Reads your codebase and writes a project README.md that doesn't read as AI-generated, with run commands checked against build files and screenshots from the project. Supports 20 project types.

Reusable AI helper assets (agents, skills, instructions, prompts, workflows) for rapid project bootstrapping.

Utility functions, types, and HTTP client for Basekit projects

Translation workbench for DeepSeek Harness: a sidebar panel that translates text through Bing or Google with retries, caching, rate limiting and automatic provider failover, plus a translate_text tool for the agent. | DeepSeek Harness 翻译插件：侧边栏翻译面板，支持 Bing

Agent skill that rewrites sloppy prompts in ASD-STE100 style (~80%) and picks the right output format (text / diagram / HTML). Inspired by Andrej Karpathy's Oct 2026 post.

Luwu (陆吾): reach the terminals and coding agents on all your computers from your phone, tablet or another computer. Secure P2P. The command is `luwu`.

Generate and edit images with OpenAI GPT Image 2.5 (or an OpenAI-compatible relay): text-to-image, image-to-image, masked edits, transparent backgrounds and custom sizes, right inside Finch.

AI Stupid Level Smart Router as a Pi provider: routing modes such as auto-coding send each request to the model with the best measured benchmark score, using your own provider keys with zero token markup.

Persistent client-facing cron automation service for DeepSeek Harness

The free workflow engine: runs, actions as steps, branching, waits, and triggers — manual, API, a schedule, a job finishing. No HTTP: @xeplr/workflow-api puts the routes on top.

Evolution-aware code search for git repositories: find current code by its former names, and ask questions grounded in code and history.

Runs coding CLIs and routes each turn to the cheapest sufficient model using TypeSafe Jev

Unified AI Tooling, Context-Aware Tool Registry, Live Partial-JSON & Stream Parser, Multi-Provider Clients (OpenAI, Claude, Gemini, Grok, LM Studio), and Live AI Chat UI

Library that lets NPCs think, plan, and act in games, powered by local LLMs.

Morphe in your terminal. Say what you want built and the agent makes it in your project folder, on whichever model you bring: Claude, GPT, Gemini, Groq, DeepSeek, xAI, OpenRouter, or a local Ollama / LM Studio model.

把 B 站 / 抖音视频整理成 Obsidian 学习笔记的 AI 技能：本地转写、面试题与学习两种模式、自动写库。

OpenCode provider plugin: Gemini on Vertex AI behind an Apigee gateway (OAuth client credentials).

LaTeX correction through your own Ollama endpoint

Inline video and audio playback for DeepSeek Harness: a play_video tool plus a native player under the reply and in the sidebar, served by a guarded, Range-capable streaming route, and a look_video tool that shows the model frame contact sheets with audio

Deferred tool loading for agents on OpenAI-compatible Chat Completions APIs.

Two Claude skills that review product design portfolios (IC and leadership) from a hiring manager's point of view.

Lint prompt template files before they ship: unfilled {{variables}}, duplicate or contradicting instructions, token budgets, broken JSON examples and unclosed XML tags. Zero dependencies.

Native text-to-speech for Cordova on Android, iOS and browser.

A free static site generator for AI website development: Markdown pages with built-in SEO, llms.txt, checks, spam protection and safe deploys.

A proactive, human-like personal-assistant kernel for DeepSeek Harness: personal knowledge, stable and situational guidance, follow-through tasks, adaptive check-ins, resource discovery, and reusable experience.