External Claude Code and Codex assets¶
Backend.AI GO can link to skills and agents on its host or import editable copies. Review external definitions before enabling them. A linked skill also needs approval of its current definition before use. Enabling an item and approving its definition do not approve the tool calls it makes.
Connect a source¶
- Open Agent Extensions, choose Connected sources, then Connect source.
- Select Claude Code or Codex and choose the actual
.claudeor.codexdirectory. For a headless server, choose a directory on the server. - Connect and review the discovered items. New linked items start disabled.
- Open a skill, review its definition, and approve that revision. Enable the items you want to use. Bulk enable changes enablement only.
A Claude Code connection reads skills/*/SKILL.md, commands/*.md, and agents/*.md. A Codex connection reads skills/*/SKILL.md, prompts/*.md, and agents/*.toml. Connecting a directory does not load its instruction files into every run. Instruction loading is a separate setting for workspace runs.
Linked, copied, and detached items¶
| Mode | Content used | How to edit or update |
|---|---|---|
| Linked | Current external definition | Edit the source file, or detach to edit in the app |
| Copied | Imported application copy | Edit in the app; re-import explicitly to refresh from the source |
| Detached | Application-owned copy | Edit in the app; it no longer follows its source |
Linked definitions are read when inspected and checked again before execution. There is no background file watcher. Detaching preserves the item's identity, enablement, and user policy while copying the current definition and supporting files. Disconnecting a source removes its linked items; copied and detached items remain.
In the skill editor, Definition edits the skill text and frontmatter. Supporting files manages scripts and other files below the skill directory. An unsaved definition stays in the editor when switching tabs. Save the definition separately. Linked skills must be detached before editing.
Supporting files are limited to 512 files, 1,024 total entries, depth 8, 16 MiB per file, and 64 MiB in total. SKILL.md belongs to the definition editor. Symbolic links and reparse traversal are rejected. Import and export preserve ordinary Unix executable permissions while stripping privileged bits.
Changes and name collisions¶
| State | Meaning |
|---|---|
| Changed | The definition differs from its previous revision. Review and approve a changed linked skill before using it again |
| Missing | The source file is unavailable; the item cannot run |
| Invalid | The source cannot be safely read or parsed; correct the source before use |
| Shadowed | Another item with the same name wins the runtime's selection |
For skills with the same name, application-owned items precede Claude Code items, which precede Codex items. The runtime does not silently fall back to a shadowed definition when the winner is unapproved or invalid. A workspace item takes precedence over a global item only within the same origin.
For a linked skill, definition approval covers the exact definition file bytes. Supporting scripts remain live and are not covered by that hash. Changing the definition invalidates approval. A skill's allowed-tools cannot bypass normal tool approval or a user's denied-tool policy.
Installed Claude Code plugins¶
Choose the installed-plugin source option to discover plugins already installed on the host. Review the offered version and connect it. Plugin skills, commands, and agents become linked items and start disabled. This does not install a new plugin in Claude Code.
After upgrading a plugin, open Connect source and connect that installed plugin again. The existing source and stable item identities are updated to the offered version; changed skill definitions need approval again. Disconnect removes the linked items but leaves the Claude Code installation in place. Plugin hooks, output styles, and declared MCP servers are skipped and shown as unsupported assets.
Compatibility¶
Tool mapping translates names into Backend.AI GO tools. It does not reproduce another application's permission system. Availability still depends on the current run, profile restrictions, and tool approvals.
| Asset or feature | Support in Backend.AI GO |
|---|---|
| Claude Code skills | Supported as skills with frontmatter and supporting files |
| Claude Code commands | Mapped to skills |
| Claude Code subagents | Mapped to agent profiles; delegation remains a separate opt-in |
| Codex skills | Supported as skills |
| Codex prompts | Mapped to skills |
| Codex TOML agents | Mapped to agent profiles; supported reasoning effort becomes a validated default |
| Installed plugin skills, commands, agents | Linked explicitly from an installed plugin; a copy of a plugin is imported from the Claude Code Plugins tab instead |
| Plugin hooks | Skipped; not executed |
| Plugin output styles | Skipped |
Read | read_file |
Write | write_file |
Edit | write_file |
MultiEdit | write_file |
Glob | search_files |
Grep | search_content |
Bash | run_shell, executed by the runtime's run_command tool |
WebFetch | fetch_url |
WebSearch | web_search |
AskUserQuestion | select_option |
Task | delegate, subject to delegation opt-in and limits |
Agent | delegate, subject to delegation opt-in and limits |
Skill | use_skill, subject to selection and revision approval |
TodoWrite | Unsupported; retained as source metadata |
NotebookEdit | Unsupported; retained as source metadata |
mcp__* | Unsupported; retained as source metadata |
Codex shell | run_shell |
Codex read_file | read_file |
Codex apply_patch | write_file |
| Source model names | Hints only; do not select a Backend.AI GO runtime model |
Argument scopes such as Bash(git *) | Retained as information; not enforced as argument allowlists |
Linked allowed-tools auto-approval | Not inherited; normal tool approvals still apply |
${CLAUDE_SKILL_DIR} | Expands to the active skill directory; host script execution requires workspace mode and shell approval |
CLAUDE.md, AGENTS.md, user rules | Explicit, default-off workspace instruction loading |
Headless use¶
A source path belongs to the machine running Backend.AI GO. ~/.claude, ~/.codex, and plugin discovery use that server process's home directory. Connecting from a browser does not upload the browser machine's files.
The server exposes the same source lifecycle and revision approval through its authenticated API. See Headless mode and the CLI reference for connection and authentication setup. A workspace workflow that needs git, gh, or scripts needs the server's installed executables and credentials, plus an approved host workspace.