<picture> <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/conn-castle/agent-layer-web/main/static/img/branding/logo-dark.svg"> <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/conn-castle/agent-layer-web/main/static/img/branding/logo.svg"> <img src="https://raw.githubusercontent.com/conn-castle/agent-layer-web/main/static/img/branding/logo.svg" alt="Agent Layer logo" width="120"> </picture>

Agent Layer

One repo-local source of truth for instructions, skills, MCP servers, and approvals across coding agents.

Agent Layer keeps AI-assisted development consistent across tools by generating each client’s required config from a single .agent-layer/ folder. Install al once per machine, run al init per repo, then run al <client> (e.g., al claude) to sync and launch.

Key properties:

• Local-first, no telemetry
• Deterministic outputs from canonical inputs
• Explicit approvals and command allowlists
• Optional per-repo runtime isolation (Codex auth/sessions/logs; Claude settings and caches with auth pending upstream fix)

Comparison:

Manual per-client setup	Agent Layer
duplicate instructions across multiple formats	one canonical source under .agent-layer/
inconsistent approvals and command policies	consistent approvals and allowlists
MCP servers added in one client and forgotten in another	generated MCP config for supported MCP clients
shared global state across repos	opt-in per-repo Codex auth/sessions/logs; opt-in Claude settings and caches isolation
no single place to review or audit changes	audit in version control

If Agent Layer improves your workflow, please consider starring the repository. Stars help new users discover the project.

---

Supported clients

MCP = Model Context Protocol (tool/data servers).

Client	Instructions	Skills	MCP servers	Approved commands
Antigravity	✅	✅	❌*	✅**
Claude Code CLI	✅	✅	✅	✅
VS Code / Copilot Chat	✅	✅	✅	✅
Codex CLI	✅	✅	✅	✅
Copilot CLI	✅	✅	✅	✅

Notes:

• Codex, Antigravity, VS Code/Copilot, and Copilot CLI consume shared skills from .agents/skills/<name>/SKILL.md.
• Claude Code consumes skills from .claude/skills/<name>/SKILL.md.
• Claude Code and Codex VS Code extension support is handled through al vscode.
• Auto-approval capabilities vary by client; approvals.mode is applied on a best-effort basis.
• *Antigravity MCP config is written to .agy/antigravity-cli/mcp_config.json using the supported serverUrl shape. agy v1.0.0 migrates that file to <gemini_dir>/config/mcp_config.json, but runtime MCP registration has not been observed yet. Run al probe agy to see the current capability matrix.
• **Antigravity approved commands are written to .agy/antigravity-cli/settings.json as a managed permissions.allow list. User passthrough in agents.antigravity.agent_specific can add other settings keys, including permissions.deny; runtime enforcement is reported by al probe agy.

---

Install

Install once per machine; choose one:

Homebrew (macOS/Linux)

brew install conn-castle/tap/agent-layer

Script (macOS/Linux)

curl -fsSL https://github.com/conn-castle/agent-layer/releases/latest/download/al-install.sh | bash

Verify:

al --version

---

Quick start

Initialize a repo (run from any subdirectory):

cd /path/to/repo
al init

Then run an agent:

al agy

Optional health check:

al doctor

Notes:

• al init prompts to run al wizard after creating the bare operational scaffold. Use al init --no-wizard to skip; non-interactive shells skip automatically.
• al init is intended to be run once per repo. If the repo is already initialized, use al upgrade plan and al upgrade to refresh template-managed files.
• By default al init first walks