One layer for every
coding agent.
Cross-agent message dispatch, portable skill sync, and one repo-local source of truth.
# Install agent-layer
$ brew install conn-castle/tap/agent-layer
# Initialize agent-layer in your repo
$ al init
# Launch your agent (sync happens automatically)
$ al codex
Featured clients
GrokMuseWhy Agent Layer?
Use one repo-local source of truth for Claude Code, Codex, Grok, Muse, Antigravity, Copilot CLI, and VS Code.
Agent Dispatch
Let an agent delegate bounded work to Claude, Codex, Grok, Muse, or Antigravity through seven MCP tools. The same handle-based lifecycle is available to people and automation through the CLI. Explore Agent Dispatch.
Git-backed Skill Imports
Import Agent Skills from any reachable Git repository, edit them locally, merge upstream changes, and contribute improvements back—without losing your project-specific work. Import Agent Skills.
Single Source of Truth
Review instructions, skills, approvals, and MCP servers as code under .agent-layer/. Agent Layer writes each client's native configuration when you sync or launch.
DeltaSelect
Design a lower-cost benchmark for instruction and skill experiments by selecting tasks most likely to detect a real effect within your budget. See DeltaSelect.
Consistent Safety and MCP
Define approval rules and MCP servers once. Agent Layer projects the closest supported behavior into each enabled client and reports capability gaps.
Local, Versioned, and Auditable
Pin Agent Layer per repo, keep team behavior reproducible, and work without product telemetry. See Security for the exact network and secret-handling boundaries.
Install Agent Layer
Install on macOS or Linux, then run al init in a repo.
If Agent Layer is useful, consider starring the repository. Stars help new users find the project.