The full catalogue, with the reasoning behind the decisions that look arbitrary. For the short
version with screenshots, see README.md.
The header has two sections: Social (Instagram) and Sales (WhatsApp conversations, contacts and everything that sells). Settings is the gear beside the bell.
Off until SOCIAL_ENCRYPTION_KEY is set; connecting an account is in
social-section.md.
Rules that answer people on your behalf. A rule has one trigger — a comment (on one post or any post, optionally only when it contains certain words), a DM keyword, a story reply, a story mention or a button tap — and then any of:
- a public reply under the comment, one of several wordings picked at random so it does not read as a bot;
- an opening message ("tap below and I'll send it") and an optional follow gate;
- the DM — text, a link card with buttons, a card, or quick replies;
- questions that collect name, email, phone or free text, saved under Leads and optionally to a Google Sheet;
- one follow-up if they don't tap the link, at a delay you pick inside Instagram's 24-hour window.
Each rule is edited step by step beside a phone that previews the post, the comment, the reply and the DM, and Test it walks a fake commenter through the whole thing. New automations start from the Auto-DM links from comments preset.
Links in a DM are tracked through a short redirect on this install, so each rule reports sends, clicks, CTR and leads. The redirect is public by design — the person tapping is not logged in.
Multi-step DM conversations on a canvas, for when one rule is not enough, with a step-by-step simulator. Like rules, a flow is created switched off.
An agent that answers DMs and/or comments for one account, from a prompt and an OpenRouter model, with an optional hand-off to a person, auto-tags and a test chat. Rules, flows and agents share one Automations page with their results and an activity log.
- Overview — followers, reach, engagement rate and top posts over 7, 30 or 90 days.
- Content library — every synced post and reel with views, engagement, rate and a content score; a Media tab for the account's images and video.
- Inbox — DMs and comments in one place.
- Contacts — the people behind them, with their own tags and fields.
All of it reads the synced database, never the Graph API on page load, so browsing never spends the account's rate budget. Every number says when it was last synced.
On Meta's WhatsApp Cloud API, one or more numbers per install. Connecting a number is in
whatsapp.md.
A left-to-right visual editor with drag-to-connect wiring. Triggers: keyword, new contact, any message and tag applied. Steps: messages (text, media, buttons, lists, call-to-action links, approved templates), conditions, actions (funnel stage, tags, fields, owner) and delays.
Every branch a step can take is a labelled row on the card with its own connector, so a reply button, a list option and a timeout are each visibly wired rather than sharing one anonymous handle.
A flow is checked before it can go live: an unwired button, an empty keyword or a step that asks something without waiting for the answer is refused with a reason. Preview & test runs it as a fake contact.
A no-code LLM agent per WhatsApp number: system prompt, model choice, triggers, multi-turn memory, a tool-use loop and a full run trace. Tools cover Google Sheets, HTTP requests, media sends and CRM write-back.
- Human handoff by keyword or agent decision, with round-robin assignment, idle-conversation summaries, voice-note transcription and vision.
- Usage limits cap what an agent may spend on one person — a rolling per-person allowance that refills on a clock rather than a lifetime cap. Test numbers are exempt from the pause, so a limit can never lock the operator out of their own rehearsal.
- Created from presets (answer questions, get leads, sell & book, support) filled in from a short form, then tuned in the editor beside a live test chat.
Agent runs execute on a queue outside the webhook path, so Meta's 20-second timeout is never hit.
Agents that work across the whole workspace rather than one chat: run on a schedule or on a trigger (a new lead, a stage change), read leads, conversations and sales within a data scope you set, and produce drafts, worklist items or actions. Anything that would message a customer or change a record waits in an approvals queue unless you allow it outright, and every run keeps its evidence.
The full Meta lifecycle in a five-step builder with a phone preview: submit, sync, edit, delete — with PAUSED / DISABLED / REJECTED handling, quality score, COPY_CODE buttons, carousels and library clone.
Editing an approved template snapshots the previous version, enforces Meta's 2-edits-per-24h limit, and offers restore from a history drawer. Cached daily analytics include a per-button click breakdown.
A 3-pane WhatsApp client:
- per-agent filtering, with numbers masked by default;
- media — image, video, audio, document;
- 24-hour customer-service-window enforcement;
- optimistic-UI sends and mic recording in the composer;
- voice notes transcribed to text (OpenRouter, multilingual).
Voice notes. A recording, a forwarded voice note and Media-library audio go out as a real WhatsApp voice note — profile picture, mic icon — in WhatsApp's own file layout: 48 kHz mono Opus in 120 ms packets, the audio on one Ogg page, vendor "WhatsApp". ffmpeg's default OGG is delivered and then "This audio is no longer available" on iPhone. Audio over the 512 KB voice-note limit goes out as an ordinary M4A audio message.
7 message types, a per-recipient queue, live SENDING / SENT / PARTIAL / FAILED rollup, and per-broadcast variable mapping. The composer asks for the send mode first — send now, schedule once, or repeat on a schedule — because every later choice depends on it. Recipients can be filtered by funnel stage, tag and date range, and a scheduled broadcast stays editable until it fires.
Upload once to MinIO, sync per-account to Meta on demand. Each account gets its own 28-day media id, with an optional daily cron that refreshes ids before they expire.
What the messaging above actually owes Meta, per template and per message type. Meta puts a pricing
object on every status webhook; this app gives it a permanent home.
The money amount is derived from the WABA's own pricing analytics (cost ÷ volume) rather than a hand-maintained rate card.
Rates differ enormously by country. Measured on a real account, India utility billed 0.1150 against Germany's 4.0322 for the same category — one hardcoded rate would have under-reported by ~97%.
Every send is stamped with its template and originating surface — broadcast, automation, agent, manual or MCP — at send time, because working out "which template was this?" afterwards silently misses.
The funnel for the period in one row — came in, reached out, got a reply, won, paid — with open conversations, active automations, where leads came from and how many sit in each stage.
One hub over one leads model — Funnel and All Contacts (table or board) are views of the same
records rather than separate pages. Tags, a configurable field registry, and CSV/XLSX import with
alias-matched headers, drag-drop, Ctrl+V paste and idempotent upsert.
Stage labels, colours, order and won-flag are all editable. The underlying
stage_keyis immutable, so renaming a stage never rewrites existing rows or breaks conversion maths.
Rename "Profession", hide "Pincode", edit a dropdown's options, or add a custom field — with no
schema change. Built-in columns are relabelled rather than replaced; field_key is immutable once
minted; deleting a custom field is a soft delete and the key stays reserved forever.
That last rule is not tidiness. A new field re-using an old key would silently inherit orphaned values still sitting in existing JSONB blobs.
Pin a lead's chat and give it a follow-up date and time — or let the AI suggest one from the chat's timing and recent messages. The calendar shows the month with each day's follow-ups, a side panel of what is overdue and due today with Open chat and Reschedule, and drag-to-reschedule; the Chats view lists pinned leads beside their conversation. With Google connected, each follow-up is mirrored as a calendar event that links back to the chat.
A sale is a lead in a won stage. Payments and refunds against it are recorded by hand
(POST /students/:id/installments) — nothing arrives from a payment gateway. Products are the
sellable catalogue, with optional default prices.
The app is itself an MCP server, so an assistant like Claude can operate it as a custom connector. Tools are grouped into categories, every category defaulting to off.
A tool belongs to exactly one category, and the category is the gate. Categories are named after the job someone is doing ("Template Builder", "Instagram automations & flows") rather than the internal route they call, and each is tagged with what it can do:
| Tier | Meaning |
|---|---|
| Reads only | Cannot change anything or reach a customer |
| Builds & configures | Creates or edits setup — templates, agents, flows, funnel stages |
| Costs money per use | Bills your AI provider per call |
| Reaches customers | Sends real messages to real people. Meta charges apply |
| Cannot be undone | Permanently removes something |
| Full API access | Unrestricted internal API calls, scoped separately by area |
Capabilities are global, not per-token. Turning one off applies immediately to every already-connected client.
What Claude builds is a draft. create_automation (WhatsApp flows), social_create_automation,
social_create_comment_dm_draft (the comment-to-DM preset in one call) and social_create_flow all
save switched off; a separate call switches them on once you have looked.
- OAuth 2.1 at
https://<your-domain>/api/mcp— the recommended transport. In Settings → Claude & MCP, paste the server URL into Claude's custom connector and sign in with your Forge Growth login. Workspaces that require a client ID and secret create one under Access & keys. - The Claude plugin, downloaded from the same page. The
.zipis built per request with this install's own address already written into.mcp.json, so there is nothing to edit. - A legacy key-in-URL transport at
/api/mcp/http/<key>, and a local stdio server (mcp-server/).
The address is worked out per request, so the same build works on any domain. Behind a proxy that
rewrites the Host header or hides https, set PUBLIC_URL=https://<your-domain> in .env and restart;
Settings → Claude & MCP shows which address is in use and where it came from. Use the address you
sign in at: the token is issued for the host the request arrived on.
Check these first when a connector will not finish authorising:
- HTTPS with a valid certificate. OAuth discovery is refused over plain HTTP.
/.well-known/oauth-authorization-serverand/.well-known/oauth-protected-resourcemust reach the backend. The bundled nginx config already proxies them; a custom reverse proxy forwarding only/apiwill serve the SPA's HTML for these paths instead, and discovery fails with no useful error.- PKCE
S256. A missing orplainchallenge is refused rather than downgraded.
OAuth connect for Sheets, Calendar and Gmail, with built-in discovery browsers — spreadsheet picker with tab preview, calendar list, Gmail labels — and an automation action for each.
Settings is grouped by what you are setting up, and each page lists its sections as pills:
| Group | Pages |
|---|---|
| Workspace | General (preferences, Instagram notifications) · Team · Activity log |
| Contacts | Tags (WhatsApp and Instagram tags and categories) · Fields · Funnel |
| Connections | WhatsApp (accounts, webhook) · Instagram (connection, privacy policy) · AI models · Google · Claude & MCP |
Roles are rows, not an enum. admin is fixed; the rest are managed under Settings → Team,
each owning its own page list — so adding a role never needs a migration. Plus per-user number
assignments, per-contact assignment overrides, and an append-only audit log.
Under Settings → WhatsApp → Webhook: every inbound payload audited with its parser outcome, a synthetic payload generator for testing, and a replay button that re-runs any historical payload through the handler.