{% extends "base.html" %} {% block title %}Native AI Agent Portal{% endblock %} {% block page_id %}native_agents{% endblock %} {% block head %}{% endblock %} {% block content %}
The desktop's saved ConfigManager.ai_agents and ConfigManager.claude_agents catalogs. Home's AgentManager templates are a separate catalog.
Save changes the native configuration on this PC; it does not activate an agent or contact a provider. Private fields remain blank here. Leave them blank to keep saved values, or explicitly clear them. Unknown native fields are preserved.
Editing this JSON retains hidden fields. Private fields must use the separate controls above. Ambiguous edits to partially redacted values are refused.
Test sends the current draft using the shared desktop AIClient and effective global defaults, with no cached reply, retries, or routing override. This contacts the actual endpoint and may incur provider charges. Validate All sends sequential real requests for the saved Claude catalog. Stop closes only this browser's owned worker; a request already received by a remote provider can still finish there.
General Test All uses the saved General catalog, skips disabled agents, and requires a saved API key before contact, matching the desktop. It sends Hello through the actual shared AIClient with native persona, global defaults and each agent's Remote Models override setting. Cache and retries are disabled; the owned job has an absolute time budget. Unsaved drafts are retained and are not tested by Test All. Results distinguish failed, missing-key, disabled and unprocessed records; unavailable cost remains unknown.
Import appends checked native records. Claude import sets provider to anthropic, as the desktop does. Export preserves the native list format with credentials removed; it is not a private-key backup.