Capability Native Ingestion

Native CC-tree ingestion

Several tools read Claude Code's own config or plugin trees directly — the exact tree agentgear's materializer already produces. Where a tool ingests the full plugin tree natively, translating its config is redundant, so agentgear retires the overlapping translation where the native path already covers it. This is the cross-cutting integration story: for some harnesses, the best integration is to let the tool read the CC tree, not to write its config.

Who reads the tree

harness	reads	scope of ingestion
codex	codex plugin marketplace add/add parse .claude-plugin/marketplace.json + plugin.json	full tree. mcp live-merged at read, hooks copied but never fire
cursor	aggregator reads ~/.claude/plugins/installed_plugins.json, then loads .claude-plugin/plugin.json per install	full tree with ${CLAUDE_PLUGIN_ROOT} substitution (source-proven, undocumented)
droid	droid plugin marketplace add reads .claude-plugin/marketplace.json + plugin.json	full tree → ~/.factory/plugins/ (binary-proven end to end)
qwen-code	qwen extensions install converts a .claude-plugin/plugin.json dir into a qwen extension	full tree, install-time and manual; carries skills through
omp	first-class claude-plugins provider reads ~/.claude/plugins/installed_plugins.json	agents only
openclaw	plugins.load.paths append, or drop the tree at ~/.openclaw/extensions/<id>/ (auto-detect)	full tree, "Claude-compatible bundles"; hooks still register nothing
jetbrains-copilot	bundled agent's marketplace service reads .claude-plugin/plugin.json, hooks/hooks.json, ${CLAUDE_PLUGIN_ROOT}	full tree per the format descriptor (source-proven, no IDE launched)
antigravity-cli	agy plugin import <path> copies the source dir verbatim	full tree, path-only. skills + agents ingest; hooks copied raw and never fire; mcpServers dropped
augment	auggie plugin marketplace add / --plugin-dir read .augment-plugin/ or .claude-plugin/	full tree (vendor-documented; not exercised past the auth wall)
vscode-copilot	VS Code core discovers .claude-plugin/marketplace.json + plugin.json, expands ${CLAUDE_PLUGIN_ROOT}	full tree, plus default-on loose .claude/ hooks, agents, skills
devin	read_config_from.claude live-merges ~/.claude.json, .claude/settings.json, ~/.claude/skills, ~/.claude/agents, CLAUDE.md	loose config only, not plugin bundles
amp	reads .claude/skills, ~/.claude/skills, and CC's plugin-cache skills (~/.claude/plugins/cache/**/skills/*)	partial: skills only, but a plugin's skills surface via the cache

Clean negatives (no plugin-tree ingestion): antigravity (the IDE, distinct from agy), cline, crush (reads loose .claude/skills as skill dirs only), gemini, goose, kilo, kimi, kiro, opencode, pi, zed.

Where translation retires

The native path is not just a curiosity — agentgear stands its translation down when the tool already covers the plugin, to avoid double-registration:

• copilot-cli goes furthest: since 2026-07-18, native ingestion is the backend's whole mechanism, not a redundant path beside a translation.
• omp's agent write retires when Claude Code's registry lists the plugin and omp's claude-plugins provider is on (default). Double-registration then only bites if the provider is user-disabled or CLAUDE_CONFIG_DIR relocates the registry off the HOME path omp reads.
• cursor's whole translation retires when Claude Code's registry (HOME-based) lists the plugin. The gate does not check cursor's enabledPlugins, a deliberately conservative gap: a CC-side-disabled-but-installed plugin still reads as covered.

Gotchas

• Ingestion scope is the honest limit. "full tree" for antigravity-cli still drops mcpServers and never fires the copied hooks; codex live-merges mcp but its copied hooks never register. Read the scope column, not just the ingests-or-not.
• Loose config ≠ plugin bundle. devin and crush read loose .claude/ files rather than a plugin-tracked tree, so a skill inside a plugin never surfaces there. Only a loose ~/.claude/skills/<name>/SKILL.md does. amp is the partial case: it reaches CC's plugin-cache skills, so a plugin's skills do surface (skills only, never mcp/hooks/commands).

See also

• Capabilities — the index.
• Agents — where the agent translation retires.
• How it works — the materialize pipeline that produces the tree tools read.