AgentSeed

One command. Safety rules + scenario packs for every AI coding agent.

pip install agentseed && agentseed forge

English · 中文 · 日本語

GitHub (primary) · Gitee · Gitcode

---

Without AgentSeed: Every new AI session = 10 min re-explaining "don't rm -rf, don't hardcode keys, here's my stack." With AgentSeed: agentseed forge — safety kernel + scenario pack, synced to 15 platforms.

How it works

┌─────────────────────────────────────────────────────────────┐
│  L4 RUNTIME      CLI (--json) · MCP Server · HTTP           │
├─────────────────────────────────────────────────────────────┤
│  L3 ADAPTERS     15 platforms · pre-tool-use hooks          │
├─────────────────────────────────────────────────────────────┤
│  L2 CAPABILITIES  research · testing · review · DAR (on-demand) │
├─────────────────────────────────────────────────────────────┤
│  L1 SCENARIO PACKS  coding · novel · paper · devops · …    │  ← pluggable
├─────────────────────────────────────────────────────────────┤
│  L0 GOVERNANCE    P0 red lines · instruction budget · fail-stop │  ← immutable
└─────────────────────────────────────────────────────────────┘

1. L0 Governance Kernel — P0 safety rules (no rm -rf, no hardcoded secrets, no unprompted installs). Immutable. No scenario pack can override it.
2. L1 Scenario Packs — coding (default), novel, paper, agent-builder, data-science, fullstack-web, devops. Mutually exclusive, auto-routed from project anchors.
3. L2 Capabilities — testing, research, review, creative. Loaded on-demand to respect instruction budget.
4. L3 Adapters — Same rule set rendered into each platform's native format + pre-tool-use interceptors.
5. L4 Runtime — CLI, MCP Server (stdio/HTTP), governance checks.

Quick start

pip install agentseed
agentseed forge

That's it. From empty directory to a 1200-line AGENTS.md with safety rules, project skills, and platform configs.

# Generate for a specific platform
agentseed apply --profile coding --tool claude-code   # → CLAUDE.md
agentseed apply --profile coding --tool cursor         # → .cursor/rules/project.mdc
agentseed apply --profile coding --tool copilot        # → .github/copilot-instructions.md

# Or sync to all 15 platforms at once
agentseed sync

# Start MCP server (for MCP-compatible clients)
agentseed serve              # stdio mode
agentseed serve --port 8080   # HTTP + SSE mode

Supported platforms

Platform	Output file	Hook
Claude Code	CLAUDE.md	✅ PreToolUse
Cursor	.cursor/rules/project.mdc	✅
GitHub Copilot	.github/copilot-instructions.md	—
Windsurf	.windsurfrules	—
Gemini	GEMINI.md	✅
Trae	.trae/rules/project_rules.md	✅
Cline	.clinerules/project.md	✅
Continue	.continue/rules/project.md	—
Amazon Q	.amazonq/rules/project.md	—
Codex	.codex/rules.md	✅
Qodo	best_practices.md	—
Lingma (通义灵码)	.lingma/rules/project.md	—
Comate (百度)	.comate/rules/project.mdr	—
QwenWork	AGENTS.md	—
AGENTS.md (universal)	AGENTS.md	—

What makes it different

	agent-rules	agents.md	Cursor Directory	ACP	AgentSeed
Multi-platform	❌ Cursor only	❌ format only	❌ Cursor only	❌	✅ 15 platforms
Governance kernel	❌	❌	❌	❌	✅ P0 immutable
Scenario packs	❌	❌	snippets	❌	✅ 8 packs
Pre-tool-use hooks	❌	❌	❌	❌	✅
MCP Server	❌	❌	❌	❌	✅
Self-evolution	❌	❌	❌	❌	✅
Active maintenance	archived	proposal	yes	unknown	✅

Install

# From PyPI (recommended)
pip install agentseed

# From GitHub release
pip install https://github.com/weed33834/AgentSeed/releases/download/v1.1.0/agentseed-1.1.0-py3-none-any.whl

# From source
git clone https://github.com/weed33834/AgentSeed.git
cd agentseed && pip install -e .

Usage

agentseed forge              # detect → assemble → generate
agentseed forge --dry-run    # preview without writing
agentseed forge --profile novel   # use a specific scenario pack

agentseed switch --profile paper  # switch scenario pack

agentseed sync               # sync to all platforms
agentseed sync --platform cursor   # sync to one platform

agentseed status             # what's assembled, what's missing

agentseed serve              # MCP server (stdio)
agentseed serve --port 8080  # MCP server (HTTP + SSE)

agentseed platform list      # list 15 built-in platforms
agentseed platform import my-ide --entry .myide/rules.md --format markdown

agentseed pack list          # market catalog
agentseed pack add novel     # install a scenario pack on demand
agentseed pack new my-scenario  # scaffold a custom pack

Project structure

core/                  governance kernel (P0 red lines — immutable)
personas/              scenario packs (coding, novel, paper, … — pluggable)
capabilities/          capability plugins (testing, research, review — on-demand)
adapters/hooks/        platform adapters: pre-tool-use interceptors
src/agentseed/         CLI, sync engine, router, forge, evolution

Hooks

Every platform with hook support gets a pre_tool_use.py that intercepts dangerous operations before execution:

• Hardcoded API keys/secrets → denied
• git add .env / git commit .env → denied
• rm -rf / / git push --force / git reset --hard → requires approval
• Large file rewrite (>100 lines) → warned

Fail-open design: if the hook crashes, the tool call goes through.

Contributing

See CONTRIBUTING.md. Short version: edit source files in core/, personas/, or capabilities/; run agentseed sync to regenerate; don't hand-edit generated files.

Tests: python -m pytest tests/ (175 passing).

License

Apache-2.0

---

If AgentSeed saved you from re-explaining rules to your AI, give it a ⭐