Inbox Triage

What it does

Inbox Triage is a connector-free agent for busy professionals. You hand it a batch of emails as data — there is no live email API and no credentials — and it classifies each message into urgent, action, info, noise, or suspicious, prioritizes it (high/medium/low, with optional VIP senders forced to high), and drafts a suggested reply for messages that need one. The classification and prioritization core is fully deterministic (same input ⇒ same output). Only the reply-drafting step uses an LLM, and even that has a deterministic template fallback, so it never fails on a missing model. Guardrails: the agent only classifies and DRAFTS — it never sends, archives, or modifies a mailbox. Phishing/spam-looking messages are flagged and never drafted. Email content is treated as untrusted data, so any embedded prompt-injection is neutralized and never executed, and PII is kept out of logs. Tools: - run_full — end-to-end triage: classify, prioritize, and draft suggested replies (LLM with template fallback). Never sends or archives. Returns the full triage JSON. - classify — pure deterministic classification + prioritization (no LLM, no drafts). Returns scored items and category counts. - list_capabilities — static capabilities: tools, categories, priorities, guardrails. - plan_inputs — interview helper returning questions, JSON schema, and a ready-to-edit example for a tool. Provide emails as an array of { from, subject, body, date? }; optionally pass vips and a tone preference. Array/object arguments may be passed as JSON or a JSON string. Missing required inputs return a structured needs_input payload (questions + schema + example) instead of erroring, and upgrade to a native elicitation form on clients that support it.

Example prompts

• Triage these 10 emails with run_full, treat boss@company.com as a VIP, and draft concise professional replies where needed.
• Run classify on this batch — no drafts, just give me the urgent/action/info/noise/suspicious split and priorities.
• Which of these messages are suspicious? Flag them and don't draft replies for those.