Reply to emails with an LLM

This is the canonical trigger workflow, built end to end: someone emails an address, a model reads the message and the documents attached to it, and the result goes back to whoever sent it. Four nodes, no code, no input parameters.

Input (no parameters)  →  LLM "extract"  →  Tool "send_email"  →  Output
                                 ↑
                   reads {{ trigger.* }} — the email itself

The email is not an input

Nothing here is typed into the Run dialog. The email arrives as the trigger object and the nodes read it directly — {{ trigger.subject }}, {{ trigger.text }}, {{ trigger.attachments }}. That’s the whole trick; the rest is ordinary workflow building. Background: Where the event goes.

Build it

1. New workflow. Call it something like Email → Extract → Reply. Drag on an Input node and leave its parameter list empty — this workflow takes nothing from the Run dialog.

2. Add an LLM node (id extract). Pick a capable text model. Leave max tokens and temperature at their defaults. Set the system prompt to:

   You extract structured, factual information from business email. Never invent values.

   and the prompt to:

   You are an assistant that reads an inbound email (and its attachments) and extracts what matters.
   
   From: {{ trigger.from_name }} <{{ trigger.from }}>
   Subject: {{ trigger.subject }}
   Date: {{ trigger.date }}
   
   --- EMAIL BODY ---
   {{ trigger.text }}
   
   {% if trigger.attachments %}--- ATTACHMENTS ---
   {% for a in trigger.attachments %}### {{ a.filename }} ({{ a.content_type }}, {{ a.size }} bytes)
   {% if a.text %}{{ a.text }}{% else %}(no extractable text — image or binary){% endif %}
   
   {% endfor %}{% endif %}--- TASK ---
   Reply in Markdown with exactly these sections:
   1. **Summary** — 2–3 sentences on what this email is about.
   2. **Key facts** — bullets: people, organisations, amounts, dates, reference numbers, deadlines. Quote exact values.
   3. **Action items** — who needs to do what (or "None").
   4. **Attachments** — one line per attachment saying what it contains (or "None").
   Be concrete. If something is missing or unclear, say so. No greeting, no sign-off.

   The {% for %} loop is plain Jinja — every attachment’s extracted text lands in the prompt, so an emailed PDF invoice is read by the model without any parsing code.

3. Add a Tool node (id reply) with server __builtin__ and tool send_email. Arguments:

   {
     "to": "{% if trigger.from != '[email protected]' %}{{ trigger.from }}{% endif %}",
     "subject": "Re: {{ trigger.subject }}",
     "body": "{{ extract.text }}\n\n---\n_Extracted automatically by Catalyst._",
     "format": "markdown"
   }

   to is the sender. The if is a small courtesy for the Send test button: its synthetic event comes from [email protected], and an empty to makes send_email fall back to your own address — so a test reply lands in your inbox instead of bouncing off a fake one.

4. Add an Output node so the run records what happened:

   from         → {{ trigger.from }}
   subject      → {{ trigger.subject }}
   attachments  → {{ trigger.attachments | length }}
   extracted    → {{ extract.text }}
   email_result → {{ reply.result }}

5. Wire Input → extract → reply → Output and Save.

Give it an inbox

1. Click Triggers in the toolbar → New Trigger → Email.

2. Optionally restrict Allowed senders (e.g. @yourcompany.com). Leave Extract text from attachments on. Create.

3. Copy the address from the row — it looks like [email protected].

Test it

• Send test on the trigger row fires a fake email through the real path. Within a few seconds you’ll see an accepted line under Deliveries, a run in History with an Email badge, and — because of the to fallback above — a reply in your own inbox.
• Then do it for real: email the address from your mail client with a PDF attached. The reply comes back to the address you sent from, and the run’s trigger output shows the attachment with its extracted text.

Run button vs. trigger

Pressing Run on this workflow fails at the extract node with “Template reference unresolved” — there’s no email, so there’s no trigger. That’s expected. Use Send test or a real email. If you want a workflow that works both ways, declare input parameters with defaults and fill them from the event with Map from event (see Mapping event data).

Variations

• Spreadsheets, properly. The model sees a capped text preview of an emailed .xlsx; for totals or joins add a Python node — the file is already at ./inputs/attachments/, so pd.read_excel(glob.glob('./inputs/attachments/*.xlsx')[0]) gives you real numbers.
• Route instead of reply. Add a Condition on {{ trigger.subject }} or on something the LLM extracted, and send different mail for invoices vs. everything else.
• Look at the picture. For emailed screenshots, put {{ trigger.attachments.0.url }} in a vision model’s Images field — the model sees the attachment instead of reading text.
• Feed a parameterised workflow. If the workflow already has inputs, keep them and map {{ trigger.subject }} / {{ trigger.attachments.0.url }} onto them in the trigger’s Map from event box; the same workflow then runs from a schedule, the Run dialog, or mail.