plugin/ is the repository source package for the portable wcode Skill and Agent Plugins manifests. It is not a runtime dependency: the wcode binary embeds these files at compile time, so users can run setup from any project and do not need a plugin/ directory in their current working directory.

The checked-in package is skill-only: mcp.json is valid and its mcpServers object is empty. Setup or an explicit export supplies the MCP profile when needed.

Install wcode with the verified platform installer, then use wcode update for future upgrades. Normal local configuration is one command:

wcode setup

Interactive setup offers Global (recommended) first and Current project second. Global setup writes only verified user-level Host config paths and asks for a local TTY confirmation before changing them. wcode setup --project forces project scope; wcode setup --dry-run previews changes.

Both scopes install the same minimal stdio command: wcode mcp-stdio. No repository path is embedded; the MCP Host working directory becomes the default Workspace. Existing servers are preserved and unknown/JSONC/YAML schemas stay manual.

The hidden agent-plugin command remains for advanced standalone package export:

# Instructions only
wcode agent-plugin --profile skill-only

# stdio profile; the Host working directory becomes the Workspace
wcode agent-plugin --profile local-stdio

# Remote endpoint; the client still owns OAuth
wcode agent-plugin \
  --profile remote-http \
  --remote-url https://current-host.example/mcp

See CONNECTIONS.md for profile behavior.

• .claude-plugin/plugin.json supports local Claude plugin development.
• .codex-plugin/plugin.json exposes the Skill and a profile-dependent MCP map.
• .zcode-plugin/plugin.json exposes the same portable Skill without guessing a project-root variable.
• marketplace.json describes this directory for compatible marketplaces.

These files are adapters around one Skill. They do not contain hooks, executable scripts, credentials, or a second copy of wcode's security rules. The package directory is repository/build input only: installed users do not need it in their current directory because the wcode binary embeds the canonical Skill and manifests used by setup/export.

Local agents should run only:

wcode mcp-stdio

The MCP Host working directory is the default Workspace. Use --workspace only as an explicit operator override; setup never needs to bake a repository path into the Host configuration.

Remote agents should prefer Streamable HTTP at /mcp with OAuth. Older clients may use /sse plus the session-specific /message endpoint; that path keeps the same OAuth, Origin, Workspace, Harness, and authorization checks.

Hardened first-party LSP servers may keep a bounded warm session by default. The Skill uses Tree-sitter/search for localization and semantic_navigation for cross-file relationships. --no-semantic disables LSP execution. Other LSP servers require local RiskyExecution trust bound to Workspace + server + current binary identity.

Repository-aware commands can still stop for local approval. The Skill never approves executable access, repository-operation trust, LSP session trust, or Full Access on the user's behalf.

Once wcode is connected, the Skill treats wcode as the primary repository tool surface for discovery, source reads, edits, commands, review, verification, and persistent Worklist progress. It avoids mixing another repository filesystem, edit, or command tool into the same task unless wcode lacks the capability or the user explicitly asks; mixed tools can bypass Workspace selection, stale SHA checks, and Worklist state.

Source is preserved exactly rather than minified for the model. Models localize with symbols first, then read the necessary body/range; one source read is bounded to at most 1,000 original lines.

If work truly needs the current user's Home, the user can press P in the TUI and confirm Full Access or explicitly launch with --full-access. Hard protected-path, filesystem-root, symlink/hard-link, no-shell, and command-policy boundaries remain enforced.