pack-Agent

August 16, 2026 · View on GitHub

Package any AI agent into a portable dsh mode (preset) repository.

The result is a GitHub repository that is itself a dsh preset: clone it to ~/.dsh/.agent-presets/<id> and pick it in the dsh mode selector.

What is this

pack_agent.py packages an AI agent — from Claude Code, Codex, opencode, aider, Cursor, an installed dsh preset, or any custom agent described by a JSON/YAML manifest — into a portable dsh preset repo containing:

Path	Content
`agent.cordis.yml`	dsh composition file (base preset + MCP + tools + skills wiring)
`preset.yml`	mode name / description shown in the dsh picker
`AGENTS.md`	merged agent instructions + memory
`skills/`	collected skills (one directory per skill)
`tools/`	bundled custom tool plugins
`assets/`	extra files you asked to include
`.env.example`	MCP token placeholders (never real secrets)
`install.sh`	one-command installer for dsh
`pack_agent.py`	the packer itself, so the packaged agent can be re-packed later

Highlights

Multi-source: dsh preset (--from-dsh), Claude Code (~/.claude.json, skills, CLAUDE.md, project .mcp.json), Codex (~/.codex/config.toml, memory, skills), opencode / aider / Cursor (--source opencode|aider|cursor), or a generic manifest
Secret-safe MCP: tokens/secrets are replaced with process.env.AGENT_MCP_* placeholders and listed in .env.example — nothing sensitive is committed
Memory redaction: --memory safe strips assignment-style secrets and inline credential-looking tokens (URLs and normal prose are left alone)
MCP filtering: --mcp-include, --mcp-exclude, --exclude-local-mcp (drop machine-local servers)
--dry-run previews exactly what would be packaged, writing nothing; --verbose shows detailed discovery
--push initializes git and publishes to GitHub (via gh when available)

Requirements

Python 3.9+
pip install pyyaml (TOML sources additionally need tomli on Python < 3.11)

Quick start

# Auto-detect dsh / Claude Code / Codex / opencode / aider / Cursor on this machine and merge everything
python3 pack_agent.py --id my-agent --name "My Agent" --description "..." --out dist/my-agent

# Package one installed dsh preset
python3 pack_agent.py --from-dsh standard --id my-agent --out dist/my-agent

# Package any agent described by a manifest
python3 pack_agent.py --manifest agent.manifest.json --out dist/my-agent

# Preview without writing anything
python3 pack_agent.py --manifest agent.manifest.json --out dist/my-agent --dry-run

# Only keep the github MCP, drop machine-local servers
python3 pack_agent.py --id my-agent --mcp-include github --exclude-local-mcp --out dist/my-agent

# Package and push to GitHub
python3 pack_agent.py --id my-agent --push --repo my-agent --public

CLI reference

Option	Description
`--id`	mode id (kebab-case directory name). Default: derived from `--out` or `my-agent`
`--name`, `--description`	display name / description in the dsh mode picker
`--out`	output directory. Default: `./dist/<id>`
`--source`	`auto` (default, merges all detected), `dsh`, `claude`, `codex`, `opencode`, `aider`, `cursor`, `generic`
`--from-dsh <id>`	use an installed dsh preset as the base (implies `--source dsh`)
`--manifest <path>`	JSON/YAML manifest describing a generic agent
`--skills [names...]`	skill names/paths to include; no values = include none. Default: all discovered
`--mcp-config <path>`	extra JSON/YAML/TOML file with MCP servers (repeatable)
`--mcp-include <name>`	only include these MCP servers (repeatable / comma-separated)
`--mcp-exclude <name>`	exclude these MCP servers (repeatable / comma-separated)
`--exclude-local-mcp`	drop machine-local MCP servers (home/local paths, localhost URLs)
`--agent-md <path>`	agent instructions file to include (repeatable)
`--memory`	`auto` (default) / `full` / `safe` / `none`
`--memory-files <path>`	extra memory files (repeatable)
`--extra-file <src:dst>`	copy an extra file/dir into `assets/<dst>` (repeatable)
`--tool <path>`	bundle a custom tool plugin into `tools/` (repeatable)
`--base-preset <id>`	dsh preset used as the base when the source is not dsh. Default: `standard`
`--dry-run`	print the plan, write nothing
`--verbose`	detailed discovery / redaction output
`--push`, `--repo`, `--public`, `--private`	git init + push to GitHub
`--yes`	skip prompts

Manifest format

{
  "id": "my-agent",
  "name": "My Agent",
  "description": "A packaged agent",
  "skills": ["/path/to/skill-dir", "skill-name-from-common-roots"],
  "mcp": {
    "github": {
      "transport": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "ghp_xxx" }
    }
  },
  "agent_md": ["/path/to/AGENTS.md"],
  "memory": {
    "mode": "safe",
    "files": ["/path/to/memory.md"]
  },
  "tools": [
    "/path/to/tool.mjs",
    {"path": "/path/to/tool-dir", "id": "my-tool", "name": "./tools/tool-dir/index.mjs", "config": {}}
  ],
  "extra_files": [
    {"src": "/path/to/file", "dst": "notes/guide.md"}
  ]
}

Installing a packaged agent

On the target machine:

git clone --depth 1 https://github.com/you/my-agent ~/.dsh/.agent-presets/my-agent

or inside the repo:

./install.sh https://github.com/you/my-agent my-agent

Then restart dsh and pick the mode. Set the environment variables listed in .env.example.

Security notes

MCP tokens/secrets never enter the repository — only process.env.AGENT_MCP_* placeholders plus a .env.example.
Memory defaults to safe mode, which also strips inline credential-looking tokens; use --memory full only when you really need everything verbatim.
Before publishing, review AGENTS.md, skills/ and assets/ for anything sensitive.