Workflow

Create a Workflow

Build a multi-step workflow that queries data, applies AI logic, and takes action (notify, update records, etc.).

For full API reference (all ctx methods, WorkflowResult, ops database schema, production execution), see .mug/docs/workflows.md. For cross-cutting API reference (ctx.params shape, error handling, data patterns), see .mug/docs/api.md.

Input

Workflow name or description: $ARGUMENTS

If no argument provided, ask the user what the workflow should do. Clarify:

• What data sources does it need? (which databases/tables)
• What logic or AI classification should it apply?
• What actions should it take? (SMS, email, Slack, update records, write back to external APIs via ctx.action())
• Does it need to write back to external systems? (requires a bidirectional connector() — see /connector skill)
• Should it run on a schedule or only manually?

Step 1 — Name and plan

Pick a kebab-case name for the workflow (e.g., invoice-followup, lead-scoring, daily-report).

Present a brief plan:

• Data queries needed
• AI classification steps (if any)
• Actions/notifications
• Estimated step count

Wait for user confirmation.

Step 2 — Write the workflow

Create the file at workflows/.ts:

import { workflow } from "@mugwork/mug";

workflow("", async (ctx) => {
  // Fetch all open items from the workspace database
  const rows = await ctx.query("SELECT ...");

  // Classify each item using AI with smart routing
  for (const row of rows) {
    const result = await ctx.ai("fast", {
      prompt: ``,
      system: "Reply with exactly one word.",
      maxTokens: 10,
    });
  }

  // Notify the relevant person about the results
  await ctx.notify.sms({ to: "...", message: "..." });

  // Return a summary of what happened
  return { summary: "..." };
}, {
  description: "",
  // webhook: true,  // or { auth: "hmac", secret: "WEBHOOK_SECRET" }
  // inbound: "sms",  // or "email" or "slack"
  // trigger: { source: "quickbooks", table: "invoices", on: "insert" },
});

Descriptions

Every workflow must have a description in the options object — a plain English sentence explaining what the workflow does and why.

Every ctx.* call and return statement must have a // comment on the line above describing what it does. These comments appear in the workspace explorer as human-readable step descriptions.

// Check for overdue invoices
const overdue = await ctx.query("SELECT * FROM invoices WHERE ...");

// Send a reminder SMS to the customer
await ctx.notify.sms({ to: inv.phone, message: "..." });

// No overdue invoices found, nothing to do
return { skipped: true };

ctx API reference

• ctx.query(sql, params?) — query unified workspace database (table names auto-resolved across sources). ctx.query("source", sql, params?) for scoped queries.
• ctx.exec(sql, params?) — write to workspace database. Always use the one-arg form for your own tables. The two-arg form ctx.exec("source", sql) is only for writing back to a synced connector source — never pass a made-up database name. Tables are auto-created at deploy time — if your workflow INSERTs into a table with a column list, the platform creates it automatically (TEXT columns). No separate CREATE TABLE needed. Schema evolution is automatic — adding columns to a CREATE TABLE IF NOT EXISTS will add them to the existing table without losing data.
• ctx.ai(model, { prompt, system, maxTokens?, routing?, billing? }) — returns { text, model, usage, routing? }
• Use tier names: "fast" (cheap), "balanced" (mid), "powerful" (best). For multi-provider config and BYOK, see the /ai skill.
• ctx.notify.email({ to, message, subject?, fromName?, cta? }) — send styled email with optional CTA button. For templates, surface links, and BYOK, use the /notify skill.
• ctx.notify.sms({ to, message }) — send SMS
• ctx.notify.slack({ to, message }) — send Slack message
• ctx.surfaceUrl(surfaceId, path?) — generate URL to a surface (dev/prod-aware). Use in notification CTAs.
• ctx.file(path) — read a file from files/ as ArrayBuffer (local in dev, R2 in production)
• ctx.fileText(path) — read a file as UTF-8 string. Use for templates, CSV data, JSON configs.
• ctx.collect(options) — create a form that collects data from users. Returns the form URL. For the full form schema walkthrough (field types, conditionals, pages, access modes), use the /form skill.
• ctx.secret(name) — read a workspace secret by name (from .mug/secrets). Throws if not found. ctx.credential(name) is an alias — both work identically in workflows. Use for external API keys, tokens, or credentials.
• ctx.waitFor(eventName, { timeout?, message? }) — pause workflow until external event. Returns { payload, type, timedOut }. See Step 5.
• ctx.waitForUrl(eventName) — generate a one-time callback URL for embedding in notifications. See Step 5.
• ctx.agent(name, { goal, context?, sessionKey?, caps? }) — invoke a custom AI agent. See the /agents skill.
• ctx.http(url, options?) — outbound HTTP request. Returns { status, headers, body, json, ok }. Throws on non-2xx by default. Options: { method?, headers?, body?, throwOnError?: false, retry?: { attempts? } | false, timeout?, sign?: { secret, header? } }. Auto-retries connection errors and 429 with exponential backoff.
• ctx.respond(body, status?) — set a custom HTTP response for webhook-triggered workflows. First call wins. Use for Slack URL verification, Twilio TwiML, etc.
• ctx.search(query, { source?, limit? }) — semantic search across synced data. Returns { score, table, primaryKey, row }[]. Requires deployed workspace.
• ctx.ask(question, { source?, limit?, model?, system? }) — full RAG: searches data, feeds to LLM, returns { answer, sources, usage }. Requires deployed workspace.

Drop files in the files/ directory and run mug push to upload them to production. Check .mug/manifest.json for what's available remotely.

Every ctx.* call is automatically logged with timing, input/output, and token usage.

Step 3 — Test locally

mug dev          # start the dev server (if not already running — auto-detects new workflows, no restart needed)
mug run    # execute the workflow

Workflows in workflows/ are auto-discovered — no import or restart needed.

Check the output:

• [+] = success, [x] = error
• Each step shows name, duration, and any errors
• AI steps show token usage

If the workflow errors, read the error message and fix. Common issues:

• Table doesn't exist yet → add ctx.exec() to create it, or sync the source first
• AI prompt returns unexpected format → adjust the prompt or add parsing
• No credentials for "" → check mug secret set and mug.json connection config
• AI not configured: missing MUG_AI service binding → workspace needs to be deployed first, or check mug dev is running

All ctx.* methods throw on failure. Use try/catch when you want to handle errors gracefully (e.g., continue processing other items if one notification fails). See .mug/docs/api.md for the full error handling reference.

Workflows can also be triggered by webhooks — see .mug/docs/api.md for webhook config.

Step 4 — (Optional) Pause for external events

Use ctx.waitFor() to pause a workflow until an external event arrives (approval, payment confirmation, reply). Zero cost while waiting.

// Send an approval request with a callback URL
const callbackUrl = await ctx.waitForUrl("approval");
await ctx.notify.email({
  to: manager,
  message: "New request needs approval.",
  cta: { label: "Approve", url: `${callbackUrl}?action=approve` },
});

// Pause until the manager clicks or 48 hours pass
const event = await ctx.waitFor("approval", { timeout: "48 hours" });

if (event.timedOut) {
  await ctx.notify.sms({ to: submitter, message: "Your request timed out." });
} else if (event.payload?.action === "approve") {
  await ctx.exec("UPDATE requests SET status = ? WHERE id = ?", ["approved", requestId]);
}

• ctx.waitForUrl(eventName) — generates a one-time callback URL. When visited (GET or POST), it delivers the event.
• ctx.waitFor(eventName, { timeout?, message? }) — pauses until the event arrives or timeout. Returns { payload, type, timedOut }.
• Callback URLs expire after 7 days.
• In local dev, waitFor resolves immediately with an empty payload for testing.

For full signatures and options, see .mug/docs/api.md.

Step 5 — View execution log

mug logs                   # dev server or auto-fallback to production
mug logs  --production     # force production logs

Shows every run with per-step details: timing, inputs, outputs, token counts. If no dev server is running, automatically fetches production logs.

Step 6 — (Optional) Add a schedule

Add a schedule option to your workflow. Use a cron string for simple patterns, or a structured object for complex ones:

// Simple: cron string
workflow("daily-report", handler, {
  schedule: "0 9 * * 1-5",  // weekdays at 9am
});

// Complex: structured schedule
workflow("board-meeting-prep", handler, {
  schedule: {
    weekday: "tuesday",
    nth: 3,                    // 3rd Tuesday of the month
    time: "09:00",
    skipHolidays: true,        // skip workspace holidays
  },
});

Cron strings (simple)

• "*/15 * * * *" — every 15 minutes
• "0 9 * * 1-5" — weekdays at 9am
• "0 0 * * *" — daily at midnight
• "0 9 * * 1" — Mondays at 9am

Schedule object (complex)

| Field | Type | Description | |-------|------|-------------| | weekday | "monday" .. "sunday" | Day of week | | nth | 1-5 or -1 | Nth occurrence of weekday in month (-1 = last) | | time | "HH:MM" | Time in 24h format | | interval | "15m", "1h", "6h" | Repeat interval | | between | ["14:00", "16:00"] | Time window for interval | | skipHolidays | true or "US" | Skip workspace holidays (true) or a specific country | | skipDates | ["2026-12-24"] | Skip specific dates | | cron | "0 9 1 * *" | Base cron (use with modifiers like skipHolidays) | | timezone | "America/New_York" | IANA timezone |

Holiday calendar

Configure workspace holidays in mug.json:

{
  "holidays": {
    "country": "US",
    "include": [
      "christmas-eve",
      "black-friday",
      "easter-monday",
      { "name": "Company Retreat", "rule": "3rd wednesday of september" }
    ],
    "exclude": ["columbus-day-indigenous-peoples-day"]
  }
}

Built-in holiday slugs: christmas-eve, black-friday, easter-monday. Custom rules use "Nth weekday of month" patterns. Exclude holidays by slug.

Step 7 — Deploy

mug validate     # type-checks workspace code and runs 40+ validation rules
mug deploy       # runs validation, bundles, and deploys to production

mug deploy runs validation automatically — TypeScript type errors appear as warnings but don't block deploy. Scheduled workflows run automatically in production.

When a workflow errors in production, the workspace owner and admins receive an email with the error details, failed step, and duration. Duplicate emails for the same workflow are suppressed for 5 minutes. Errors are also visible in the explorer (mug dev → Errors tab).

To trigger manually in production:

mug run  --production
mug status  

Inbound message routing

Receive SMS replies, inbound emails, and Slack interactions by routing them to a workflow. Configure via workflow options:

workflow("handle-sms", handler, { inbound: "sms" });
workflow("handle-email", handler, { inbound: "email" });
workflow("handle-slack", handler, { inbound: "slack" });

Each inbound channel delivers different ctx.params:

• SMS: { from: "+1234567890", body: "Yes, approved" }
• Email: { from: "user@example.com", subject: "Re: Invoice #42", body: "Looks good" }
• Slack: { userId: "U12345", actionId: "approve_btn", actionValue: "yes" }

Webhook URLs are shown after mug deploy: https://api.mug.work/inbound/sms/, etc.

Send-and-wait pattern: Workflow 1 sends a notification and sets a status field in the database. Workflow 2 (the inbound handler) catches the reply and checks the status field to determine what to do.

For full inbound routing reference, see .mug/docs/api.md.

---

For collecting data from users (forms with submissions), see the /form skill. For displaying data and approval UIs (list+detail pages with action buttons), see the /portal skill. For email/SMS notifications with CTA buttons and surface links, see the /notify skill.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mugwork
• Source: mugwork/mug
• License: MIT
• Homepage: https://mug.work

Install and usage instructions live in the source repository linked above.