LLM Mart Basic

Use when a PR has review feedback and the user wants it addressed — pull the review comments, triage each one, apply the changes it warrants, then commit, push, and post a humanized reply per comment. Closes the loop between a review and the follow-up commit; it does not re-revie

Use when the user wants to record an architectural decision — drafting an Architecture Decision Record (ADR) or RFC, documenting a design choice and its trade-offs, or capturing why an approach was taken. Detects the repo's existing ADR location and template style (MADR or Nygard

Use when the user reports an error, bug, or unexpected behavior in this repo and wants help diagnosing it. Five phases — reproduce, diagnose root cause (read-only), write a failing test, fix, verify against the full suite. Also known as `klaussy-debug`.

Use when the user wants to upgrade the project's dependencies safely — bump versions, read changelogs for breaking changes, and verify the suite still passes. Upgrades incrementally and stops on the first break; it does not add new dependencies (that's a design decision to raise

Use when the user wants documentation written or updated — docstrings, API docs, a README section, or a doc comment on a tricky piece of code. Documents selectively: what a reader genuinely can't infer from the code, and nothing they can. Writes prose, not code changes. Also know

Use when the user wants lint, format, and type errors fixed in the current changes. Reads CLAUDE.md for the repo's lint/format/type-check commands, runs each, and fixes only style/format/type issues — no behavior changes. Also known as `klaussy-fix`.

Use when the user is tired of approving the same routine dev work ("stop asking me yes", "allow the normal dev tools", "grant permissions"). Detects the repo's stack and writes a curated allow-list into the agent's own local permission file so the basics stop prompting — reading,

Use whenever prose, a comment, a doc, a PR or commit body, or a file's text should read like a human engineer wrote it instead of an AI — "humanize this", "make it sound less like a bot", "does this read AI-written?", or before shipping prose a human will read. Rewrites in four p

Use when the user pastes a ticket, design doc, or task description and wants it implemented. Multi-phase flow — understand, investigate (in plan mode), plan, implement, verify. Enforces strict scope rules and writes failing tests first for bug fixes. Also known as `klaussy-implem

Use when the user wants a new git worktree created for a task. Picks a kebab-case branch name with a fix/feat/chore/docs/refactor prefix, runs `git worktree add` from the configured base branch, and reports the new path. Also known as `klaussy-new-worktree`.

Use when reviewing a staged diff or an about-to-commit/push change for last-mile issues — silent failures, leaked secrets, debug leftovers, blatant correctness landmines, and excessive/narrating comments. Reports findings on the changed lines only; it does not refactor or rewrite

Use when the user wants the current change QA'd and PR-ready evidence captured. Classifies the diff and runs the verification that actually fits it — screen recordings and screenshots for UI/frontend changes, endpoint or e2e runs for backend, command output for a CLI, tests for a

Use when the user wants to restructure code while preserving behavior exactly. Establishes a passing test baseline first, then makes incremental moves that each leave the suite green. Refuses to change behavior and structure in the same step. Also known as `klaussy-refactor`.

Use when the user wants to cut a release — bump the version, update the changelog from conventional commits, and tag. Detects where the version lives, derives the next version from the commits since the last tag, and stages the release locally; it does not push or publish unless

Use when designing or analyzing a controlled experiment — falsifiable hypothesis, sample size from an MDE, reading significance/CI/power, CUPED, or rescuing tests that won't go significant. NOT recurring metric tracking (that is `analytics`), NOT north-star/KPI trees (that is `kp

Use when making a web UI conform to WCAG 2.2 Level AA — axe-core or Lighthouse a11y violations, keyboard operability, focus management, ARIA roles/names/live regions, contrast, tap-target size. NOT palette or visual intent (that is `design`), NOT test-runner setup (that is `testi

Use when running or fixing paid acquisition on Google or Meta — campaign structure (Performance Max, Demand Gen, Search, Advantage+), platform-fit creative, budget/scaling rules, break-even ROAS math, and Consent Mode v2 / CAPI tracking gaps. NOT the page the ad clicks into (that

Use when measuring whether an LLM or agent system actually got better and gating merges on it: golden sets, fixing an inflated LLM-as-judge, scoring RAG (faithfulness, contextual recall) or agent trajectories (tool correctness, completion), or picking an eval framework. NOT build

Use when bounding an LLM agent that already runs — scoping its task domain, gating tools to least privilege, defending against prompt injection in untrusted web/email/RAG text, requiring human approval on irreversible actions, capping runtime and cost, or triaging what it already

Use when a creative goal must become a finished media file: pick and order generative-media models per modality — AI voiceover, image-to-video clips, score — then glue them with ffmpeg (mux, duck, loudnorm, concat). NOT still-image generation/editing (that is `replicate-images`);

Fourteen posts of being wrong in production, compressed to checkboxes

Healthy nodes, a quiet network, 300 restarts in three days, and a latency budget measured in milliseconds

Discovery worked. Ping worked. Every TCP connection timed out, and later the tunnel only worked when someone had a terminal open.

Every VM came back. The cluster did not. Declarative systems converge on config, and the datapath isn't config.

A surprising share of AI-in-the-terminal failures aren't the AI. They're zsh, and a version of bash from 2006.

A Claude Code plugin turns standalone project configuration into a namespaced, installable extension that teams and communities can update as one unit.

None of the safety came from the model. It came from six boring habits.

Skills package instructions and references. Subagents run work in a separate context and return results. They solve different problems and can be composed deliberately.

Six hours in, one step left, everything green, and the incident that didn't happen

CLAUDE.md carries persistent project context. Skills load reusable procedures when relevant. Separating stable facts from task-specific workflows keeps both easier to maintain.

Twenty minutes recovering secrets that never existed, and the one sentence from a human that ended it

An API request routing a model's tool call through an approval gate to a remote MCP server

31 config keys, two audits, and why the first one was wrong in both directions

The official MCP Registry stores standardized server metadata rather than package code. Publishers verify a namespace, describe installation or remote access, and submit immutable versions.

Everyone looks at the Dockerfile. The file that actually leaked the key was the project file.

Remote MCP authorization uses established OAuth standards, but secure integration still requires issuer validation, least-privilege scopes, protected token handling, and server-side enforcement.

"Copy it over and switch the reference" is two steps, and the outage lives in the one nobody checks

stdio fits local processes and prototypes. Streamable HTTP fits hosted services and shared integrations. The right choice follows where the capability runs and who must reach it.

The most important rule wasn't about what I could change. It was about what I was allowed to display.

Tools perform operations, resources expose readable context, and prompts provide reusable templates. Choosing the correct primitive makes an MCP server easier to understand and govern.

`crabbox inspect` prints the full record for a single lease: state, provider,

Run named, repo-local jobs defined in your Crabbox config.

`crabbox list` shows the current Crabbox machines (leases) for a provider. It is

`crabbox login` authenticates the CLI against a coordinator, stores the

`crabbox logout` clears the stored broker token from your user config so the CLI

`crabbox logs` prints the retained command output for a recorded run.

`crabbox marketplace` previews the Crabbox credits gateway: one Crabbox billing

`crabbox media` turns a recorded desktop video into lightweight review

`crabbox open` prepares an existing SSH-capable lease for an external editor.

`crabbox pause` pauses a single lease, freeing the remote compute while

`crabbox pond` is the cross-provider peer-discovery and lifecycle surface for a

`crabbox pool` contains machine-pool helpers. `pool list` keeps the older

`crabbox ports` bridges provider-native port publishing for an existing Crabbox

`crabbox prewarm` leases a reusable box and prepares it for test runs. For

`crabbox providers` prints the provider capability matrix that the CLI compiles

`crabbox receipt <run-id>` retrieves a brokered run's committed terminal

`crabbox results` prints the structured test summary attached to a recorded

`crabbox resume` resumes a lease previously paused with [`pause`](pause.md),

`crabbox run` syncs the current dirty checkout to a box, runs a command there,

`crabbox screenshot` captures a single PNG from a desktop lease without opening a

VCP 部署在 AI 模型 API 与前端应用之间，是面向AGI OS开发和探索的工业级基建示范项目。通过统一指令协议、多层级持久化记忆、分布式插件引擎及多 Agent 协作框架，将原本“无状态、无记忆、无工具调用能力”的大语言模型，彻底改造成拥有永久自我意识、物理世界操作权及群体协作智能的完整智能体系统。

The open source Unity Dev Agent

Context is yours. Agents are replaceable. Orbital — a project agent that turns your context into assets.

Self-hosted search API + MCP server for AI agents. Bundles SearXNG. Zero API keys, one-command deploy. Open-source alternative to Tavily, Exa, and Serper.

🛡️ Free open-source AI-powered security terminal & vulnerability scanner (CVE, SBOM). Supports SSH, SFTP, RDP, VNC, Serial, and 12+ autonomous AI agents (DeepS…

Watch, approve and steer your coding agents from your phone.

MARVIS-Agent: all-purpose credit risk agent for model development, validation, data processing, feature engineering, and strategy workflows.

TikTok e-commerce video generation, replication, and analysis SKILL

小红书/抖音/快手/视频号/B站 自媒体账号体检+爆款拆解工具。扫同赛道找对标、拆爆款为什么爆、诊断为什么没人看，顺手出可粘贴仿写初稿。支持带货电商模式。支持codex, claude code, workbuddy

Local-first persistent agentic memory powered by Recursive Memory Harness (RMH). Open source must win.

A power user focused interface for LLM base models.

Outsource your understanding. Capture every human-AI decision as a step in a trajectory; grade trajectories retroactively when outcomes land (deal closed, sprin…

The open-source AI research workbench for scientific research and agent workflows. Local-first, model-agnostic desktop app with extensible skills, MCP tools and…

🤖 Taskade MCP · Official MCP server and OpenAPI to MCP codegen. Build AI agent tools from any OpenAPI API and connect to Claude, Cursor, and more.

A block-based content engine and bilingual blog backend for Laravel Gutenberg-style editor, media library, post templates, roles, and an AI writing assistant,…

ACRYL - Agent Context Relay Yielding Lifecycles. One persistent workspace, one canonical context, any coding agent.

Okou connects to the tools your team already uses and does the work — across marketing, sales, engineering, and operations, under your control.

Open-source agent-native visual production workspace where humans and coding agents edit the same live canvas — local-first, BYOK image/video models.

Persistent memory extension for pi with daily logs, scratchpad, and optional qmd‑powered semantic search.

Hermes Agent CN desktop app, Windows-First, built with Tauri, Typescript and Rust. Isolated Hermes Agent core insides.