AgentSeed

<p align="center"> <img src="assets/logo-banner.jpg" alt="AgentSeed — the governance layer for AI agents" width="100%"/> </p> <!-- mcp-name: io.github.weed33834/AgentSeed -->

pip install https://github.com/weed33834/AgentSeed/releases/download/v1.1.0/agentseed-1.1.0-py3-none-any.whl && agentseed forge

🌐 English · 中文 · 日本語

📦 GitHub (primary) · Gitee · Gitcode

---

You know the drill. Every time you open a new AI coding session, you spend the first 10 minutes reminding it not to hallucinate, not to run rm -rf, and what stack you're using. AgentSeed turns that into a non-negotiable constraint layer, then plugs in the scenario pack for your task context — once, and synced everywhere.

One command detects your project scenario, plugs in the matching scenario pack, and generates highly-constrained rule files for all your tools — Claude Code, Cursor, Copilot, Windsurf, Trae, whatever you use.

agentseed forge

That's it. From empty directory to a 1200-line AGENTS.md with safety rules, project-specific skills, and platform configs. Works the same whether you're coding, writing a novel, drafting a paper, or building another agent.

---

What you get

A baseline that doesn't get overwritten. Core safety rules (don't rm -rf, don't hallucinate credentials, don't install things unprompted) are part of the governance kernel and ship with AgentSeed — no scenario pack can override them.

Pluggable scenario packs. Kernel default mode + four starter packs: default (kernel-general, built-in), coding (your dev default), novel, paper, agent-builder. Each pack bundles a scenario protocol, prompts, skills, and a capability whitelist, routed automatically from project anchors and user intent (or switched manually with agentseed switch --profile novel). Packs are pluggable and extensible — a new scenario is just a directory plus a manifest; the kernel never changes.

All your tools, one sync. Generates the right format for 15 platforms:

• Claude Code → CLAUDE.md
• Cursor → .cursor/rules/project.mdc
• 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
• Qodo → best_practices.md
• Lingma → .lingma/rules/project.md
• Comate → .comate/rules/project.mdr
• Codex → .codex/rules.md
• QwenWork → AGENTS.md (read natively)
• AGENTS.md (works with 20+ tools that natively read it)

Hooks with teeth. Every platform gets a pre_tool_use.py interceptor that blocks dangerous operations before they execute. Fail-open design: if the hook crashes, the tool call goes through.

MCP Server. Run agentseed serve and any MCP-compatible client gets governance_check (P0 red-line check), persona_list (scenario packs), persona_activate (activate a pack), and gap_detect.

Self-evolution. AgentSeed scores capability gaps (missing tools, unknown domains) and suggests what to install. Not magic — just a weighted formula that gets better as you add more skills.

---

Install

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

Or build from source:

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

---

Usage

agentseed forge              # detect project → assemble → generate
agentseed forge --dry-run    # preview what would be generated
agentseed forge --profile coding
agentseed forge --profile novel

agentseed switch --profile paper

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

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

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

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

agentseed pack list          # market catalog (installed status)
agentseed pack add novel     # install a single scenario pack on demand
agentseed pack new my-scenario  # scaffold a custom scenario pack

agentseed persona list       # list scenario packs (command name kept for compatibility)
agentseed persona search "product manager"

---

Add your own platform

agentseed platform import my-editor --entry .myeditor/rules.md --format markdown --hook-dir .myeditor

This registers the platform, generates a pre-tool-use hook, and includes it in every agentseed sync.

---

Project structure

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

---

vs. similar projects

• agent-rules (steipete) — archived, coding-only Cursor rules.
• agents.md — file format proposal; no content, no toolchain.
• ACP — agent config manager; no governance or self-evolution.
• Cursor Directory / cursor.directory — community rule snippets; no multi-platform sync.
• AgentSeed — governance kernel + pluggable scenario packs + 15-platform sync + hooks + self-evolution, all from one CLI.

---

Contributing

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

Tests: python -m pytest tests/ (171 passing).

---

Apache-2.0