Reading your message, checking tools and knowledge, then proposing safe actions.
Classroom demo: the LLM proposes, PHP validates, MySQL remembers. Use the panels below to chat, edit tools and prompts, and watch decisions become tasks, reports, or pending approvals.
Suggested examples (click to load):
Thinking and applying safe actions…
Paste course notes or load the sample knowledge base, then ask a RAG question (see suggestions: “RAG test”).
No knowledge documents yet.
No tasks yet. They appear after the agent executes create_task.
External or high-risk tools wait for human approval (for example draft_email). Click Approve to execute after review — email drafts are stored, never sent automatically.
No pending approvals.
Tool descriptions shape when the model thinks a capability should be used. PHP still blocks disabled tools at execution time (even if the model proposes them).
No tools loaded. Check /api/tools.php.
System prompt = identity and mission. Developer prompt = format, process, and constraints. Edits are stored in MySQL only — they never edit PHP files or bypass backend validation.
No reports yet. generate_report_html stores JSON; PHP renders HTML.
Makes invisible AI behavior visible: proposed tool, confidence, teaching note, and action status.
Send a chat message after the agent is wired (steps 06–07) to see the decision JSON.
In: · Out: · Total: · Est. USD: $