Keep your AI coding tools in sync — skills, MCP servers, and compound-engineering across Claude Code, OpenCode, Codex, and Gemini CLI from a single source of truth.

$ synaxis --full
✓  Skills: 64 synced
✓  MCP: wrote OpenCode config
✓  MCP: wrote Codex config block
✓  CE: compound-engineering installed to codex
✓  CE: compound-engineering installed to opencode
✓  CE: compound-engineering installed to gemini
Done. Restart Codex/OpenCode to pick up changes.

Gemini CLI has no skills directory — CE only.

cargo install synaxis

Or build from source:

git clone https://github.com/terry-li-hm/synaxis
cd synaxis
cargo build --release
# add ./target/release/synaxis to your PATH

synaxis              # Sync skills only (fast, ~50ms)
synaxis --full       # Skills + MCP + compound-engineering
synaxis --check      # Dry run — list skills, no changes
synaxis --init       # Create default config at ~/.config/synaxis/config.toml
synaxis --help

Skills-only is the default because it's fast enough to fire on every git commit. Use --full after changing MCP config or updating compound-engineering.

Wire synaxis into your skills repo so symlinks stay fresh automatically:

# .git/hooks/post-commit
#!/bin/bash
synaxis

chmod +x .git/hooks/post-commit

Run synaxis --init to create ~/.config/synaxis/config.toml with defaults:

[skills]
source = "~/skills"
targets = [
    "~/.claude/skills",
    "~/.opencode/skills",
    "~/.codex/skills",
    "~/.agents/skills",
]
skip = ["TEMPLATE.md", ".git", ".archive", ".cache", ".gitignore", "config.json"]

[mcp]
source = "~/agent-config/mcp-servers.json"
opencode = "~/.opencode/mcp.json"
codex = "~/.codex/config.toml"

[ce]
source = "~/.claude/plugins/marketplaces/every-marketplace"
platforms = ["codex", "opencode", "gemini"]

All fields are optional — synaxis falls back to the defaults above if the config file is absent or a field is missing.

mcp-servers.json uses a split key design: mcpServers propagates to OpenCode (JSON), while _codexExtras propagates to Codex (TOML). This lets you keep OpenCode's HTTP-based servers separate from Codex's URL-based entries without duplicating config.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path"]
    }
  },
  "_codexExtras": {
    "context7": { "url": "https://mcp.context7.com/mcp" }
  }
}

The Codex TOML block is written between # synaxis-mcp-begin / # synaxis-mcp-end markers, so it coexists safely with manually maintained config.

A directory under source is treated as a skill if it contains a SKILL.md file. Symlinks and entries in skip are ignored. Each skill is symlinked by name into every target directory.

Codex gotcha: dir-level symlinks break skill discovery in Codex (#11314). synaxis works around this by creating per-skill symlinks inside a real directory rather than a single directory-level symlink.

MIT