Agents Guided Evolution
One directory makes anyproject agent-ready
AGENTS.md says what to do. .agents/ remembers what the project learned.
The problem
8 config formats. All drifting. None learning.
Projects accumulate knowledge every session. The next agent never sees it.
Every new session starts from zero. Context dies with the chat window.
CLAUDE.md, .cursorrules, GEMINI.md, copilot-instructions — all saying different things.
82% of devs use AI weekly. 59% juggle 3+ tools. Each one forgets everything.
The convention
One versioned directory. Every agent inherits it.
Rules, knowledge, and captured decisions — in git, not in chat history.
Run npx agentsge init — the project gets a structured .agents/ directory with config, rules, and a capture pipeline.
Knowledge compounds over sessions. Architecture decisions, patterns, lessons, conventions — versioned in git, loaded automatically.
The project carries its own memory. Switch agents anytime.
.agents/
├── config.yaml # stack, capabilities, metadata
├── rules/ # mandatory agent instructions
│ └── _capture.md # built-in: auto knowledge capture
├── knowledge/ # accumulated project intelligence
│ ├── _index.md # always loaded into context
│ ├── architecture/ # decisions & trade-offs
│ ├── patterns/ # repeating codebase patterns
│ ├── lessons/ # bug investigations
│ ├── conventions/ # team rules
│ └── dependencies/ # why X, known issues
├── skills/ # reusable multi-step workflows
└── mcp/ # MCP server definitions
└── config.yaml # synced to all agent configsHow it works
Three steps. Zero new workflow.
The project starts remembering. You stop re-briefing.
Initialize
npx agentsge initDetects stack and capabilities, creates .agents/ with config, capture rules, and generates AGENTS.md as the entrypoint.
Onboard with any agent
Any agent reads AGENTS.md, scans the codebase, asks targeted questions, and fills .agents/ with real project knowledge.
Knowledge compounds
Capture hooks extract decisions and lessons after each session. Every future agent inherits the accumulated context.
Convention, not product
Static instructions rot within hours. Continuous capture compounds.
Other tools configure one agent. .agents/ gives the project a memory that works across all of them.
| agents.ge | Others | |
|---|---|---|
| One source of truth | .agents/ | Multiple scattered files |
| Automatic knowledge capture | Manual or external | |
| Works across agents | Claude, Cursor, Codex, Copilot + | Single-tool only |
| MCP sync | One config, all formats | Manual per-tool |
| Existing projects | Zero disruption | Often requires rework |
| Project carries its brain | ||
| Knowledge compounds over time |
Capabilities
What lives inside .agents/
Structured layers that make project memory durable and tool-agnostic.
Automatic knowledge capture
Hooks extract decisions and lessons after each session. Pending items go through review before entering the knowledge base.
Structured project memory
Architecture decisions, patterns, lessons, conventions, and dependency notes — organized and versioned in git.
MCP sync
Define MCP servers once in .agents/mcp/, sync to Claude, Cursor, Codex, and Copilot configs with one command.
Capabilities detection
Detects what the project does — API, database, auth, AI. One config any agent understands.
Context injection
Three levels of recall — entrypoint digest, session-start hook, and per-prompt injection. The agent always has context.
Zero vendor lock-in
Plain markdown and YAML in git. Switch agents anytime — the project carries its own brain.
FAQ
Search engines and agents need the same thing: clear context.
These are the questions developers, repository visitors, and LLM-based search tools need answered fast.
What is agentsge?
agentsge is an open-source CLI that makes any repository agent-ready. It creates a versioned .agents directory with rules, project knowledge, MCP config, and capture workflows that work across multiple AI coding agents.
How is .agents different from AGENTS.md, CLAUDE.md, or .cursorrules?
AGENTS.md and tool-specific files are entrypoints. The .agents directory is the durable source of truth that stores project memory, reusable rules, structured knowledge, and MCP definitions, then syncs that context back out to each agent format.
Which AI coding agents does it work with?
agentsge is designed for mixed-agent teams and supports workflows around Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI, and other tools that can read markdown instructions or generated MCP configs.
Why does this help search and LLM discoverability?
Clear, static documentation, typed knowledge, route-level metadata, and machine-readable artifacts such as llms.txt make the project easier for search engines, repository visitors, and LLM-based search systems to understand and cite.
Start now