npx claudepluginhub bitsky-tech/amphiloop --plugin AmphiLoopBuild agents with the Bridgic Amphibious dual-mode framework — combining LLM-driven (agent) and deterministic (workflow) execution with automatic fallback and human-in-the-loop support. Use when: (1) writing code that imports from bridgic.amphibious, (2) creating AmphibiousAutoma subclasses, (3) defining CognitiveWorker think units, (4) implementing on_agent/on_workflow methods, (5) working with CognitiveContext, Exposure system, or cognitive policies, (6) adding human-in-the-loop interactions (HumanCall, request_human, request_human_tool), (7) scaffolding a new amphibious project via CLI, (8) any task involving the bridgic-amphibious framework.
Use for any task requiring a real browser: viewing web pages, accessing login-gated sites, operating web UIs, scraping social media (Xiaohongshu/Weibo/Twitter/X, etc.), reading JS-rendered or dynamic pages, bypassing bot detection, form filling, e2e checks, and general web automation. Prefer this over WebFetch whenever the page needs JS execution, authenticated session, interaction, or stealth. Invoke via terminal CLI (`bridgic-browser ...`) or Python SDK (`from bridgic.browser.session import Browser`, `from bridgic.browser.tools import BrowserToolSetBuilder`). Also covers accessibility snapshot refs, CLI-SDK mapping/migration, and generating SDK code from CLI action steps.
LLM provider initialization for bridgic projects. Use when: (1) initializing OpenAILlm, OpenAILikeLlm, or VllmServerLlm, (2) configuring OpenAIConfiguration (model, temperature, max_tokens, timeout), (3) choosing the right provider package for a task, (4) using chat/stream interfaces or advanced protocols (StructuredOutput, ToolSelection).
Multi-agent workflow orchestration with natural language creation, parallel execution, conditional flows, and visual progress tracking
Browser automation CLI for AI agents. Navigate pages, fill forms, click buttons, take screenshots, and test web apps.
A curated collection of AI coding agent skills for browser automation, frontend design, performance auditing, task tracking, and structured problem-solving workflows.
Claude Code skill pack for Apify (18 skills)
Universal Self-Learning Browser Agent — record browser interactions, generate API skills, and replay them via MCP tools with progressive tier optimization.
HelloAGENTS — The orchestration kernel that makes any AI CLI smarter. Adds intelligent routing, quality verification (Ralph Loop), safety guards, and notifications.
Uses power tools
Uses Bash, Write, or Edit tools
Share bugs, ideas, or general feedback.
English | 中文
AmphiLoop, short for Amphibious Loop, is a new methodology, technology stack, and toolchain for building AI agents. It enables tasks to be described and orchestrated using natural language, with an “Explore → Code → Verify” loop guiding code generation and build. The resulting artifacts are capable of automatically switching between workflow mode and agent mode at runtime.
AmphiLoop packages domain knowledge and execution methodology into three layers:
| Layer | Role | Description |
|---|---|---|
| Skills | Domain knowledge | "What it is, how to use it" — reference docs loaded on-demand |
| Agents | Execution methodology | "How to do it well" — specialized execution experts |
| Commands | Orchestration | Multi-step workflows that coordinate agents and skills |
Together, they enable an end-to-end pipeline: explore a website via CLI -> generate a dual-mode agent project -> verify execution — all within an agent.
Long-form pieces on the motivation and design philosophy behind AmphiLoop:
Beyond Autonomous: Why I'm Building an Amphibious Agent
# Step 1: Register the marketplace (one-time)
claude plugin marketplace add bitsky-tech/AmphiLoop
# Step 2: Install the plugin
claude plugin install AmphiLoop
Or install from a local checkout (point marketplace add at the local directory — it's read as a marketplace because the repo ships .claude-plugin/marketplace.json):
git clone https://github.com/bitsky-tech/AmphiLoop.git
claude plugin marketplace add /path/to/AmphiLoop
claude plugin install AmphiLoop
After installation, skills, agents, and commands (e.g. /build) are automatically available in Claude Code.
Commands are user-invocable workflows. Invoke them with the / prefix:
/AmphiLoop:buildUnified pipeline. Describe any task, list the domain references the agents should read (SKILLs, CLI help, SDK docs, style guides), and ask to generate a runnable project:
/AmphiLoop:build
I want to aggregate all `orders_*.csv` files under ~/data/inputs into a single
summary.csv — one row per customer with totals.
Domain flag (optional) — append --<domain> to inject pre-distilled domain context from domain-context/<domain>/. Currently supported: --browser.
/AmphiLoop:build --browser
Go to https://example.com, search for "product", and extract the first 5 results.
I want a project that can run this reliably.
Without a flag, /build auto-detects the domain from TASK.md (and falls back to a generic flow if none matches). Users can always supply additional domain references in TASK.md.
What happens under the hood:
TASK.md template where you fill in goal, expected output, and Domain References; auto-detects the domain if no flag was given--browser is active)uv, runs uv initamphibious-explore agent, which reads your domain references and probes the environmentamphibious-code agent to produce a complete project with all source filesamphibious-verify agent to inject debug instrumentation, run the project, and validate resultsAgents are execution specialists delegated by commands. They are not called directly by users but are orchestrated internally:
| Agent | What It Does |
|---|---|
| amphibious-explore | Systematically explores a target environment via a domain-supplied toolset, produces an executable plan with stability-annotated operations and supporting snapshots |
| amphibious-code | Generates a complete bridgic-amphibious project from a task description and exploration report |
| amphibious-verify | Injects debug instrumentation, runs the project with monitoring, validates results, and cleans up |
Skills are domain knowledge references that agents and Claude load automatically when relevant. You don't invoke them directly — they activate based on conversation context.
| Skill | Activates When |
|---|---|
| bridgic-browser | Using browser automation via CLI (bridgic-browser ...) or Python SDK (from bridgic.browser) |
| bridgic-amphibious | Building dual-mode agents with AmphibiousAutoma, CognitiveWorker, on_agent/on_workflow |
| bridgic-llms | Initializing LLM providers (OpenAILlm, OpenAILikeLlm, VllmServerLlm) |
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge.
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge.
Sign in to claim