FLOW AI · Agents · MCP · Skills
What is built, what is tested and what is not connected yet. Every piece states its real status, and the examples use the exact tool names.
AI assistant: built, not switched on yet MCP: in preparation, not connected
Status as of 29 September 2026. “Connected” means a real assistant completed the verification sequence with a real FLOW account; today none has.
| Piece | Built | Tested | In production | Connected or active |
|---|---|---|---|---|
| AI assistant in the appCaptions, rewriting, hashtags, ideas, alt text and replies. | Yes | LocallyAutomated tests | Not switched onThe AI provider is not configured yet. | InactiveToday it doesn’t generate suggestions. |
| ChatGPTPrivate app in developer mode · OAuth “Connect FLOW”. | Yes | Locally | Not deployed | Not connected |
| Claude on the web, desktop and mobileCustom connector · OAuth “Connect FLOW”. | Yes | Locally | Not deployed | Not connected |
| Claude CodeOAuth through the browser, or an API key. | Yes | Locally | Not deployed | Not connected |
| Claude Desktop in local modeFLOW’s local client and an API key. | Yes | Locally | Not applicable: it runs on your machine. | Not connected |
| Grok Bot (the “Dana” agent)FLOW’s client and an API key as a Secret. | YesGuide, instructions and commands. | LocallyWith FLOW’s client, not with a real Grok. | Not deployed | Not connected |
| Image and video generation (Super Flow AI)See the Super Flow AI page. | Yes | LocallyThe cost is shown before generating. | Off, no date | No generation provider is active. |
| Packaged skills for agentsReusable instructions an agent loads. | There are none | — | — | — |
“Locally” means: automated tests, a real server on a development machine and a browser. It is neither a deployed test environment nor production.
In the app
It helps you write in your brand’s voice. A person always reviews what it suggests: nothing is published on its own.
Today it generates no suggestions
MCP (Model Context Protocol) is an open standard that lets an AI assistant use another application’s tools. FLOW implements it so ChatGPT, Claude or Grok Bot can look up your brands and save drafts, always with your permission.
Automated tests that simulate the requests each assistant documents, the official MCP SDK client, a browser and FLOW’s command-line client.
The MCP server isn’t in production yet, so no assistant can connect today. It still has to be deployed and checked with each real assistant.
We haven’t checked that a real ChatGPT, Claude or Grok accepts FLOW. Until the sequence is completed with a real account, no assistant is declared connected.
They are the only ones a connected assistant will have. The names and fields come from the server code and its tests; once an assistant connects, its tool list will show these seven.
| Tool | What it does | Permission | Input | Returns |
|---|---|---|---|---|
list_brands | The workspace’s brands the connection can see. | Read | None. | brands[]: id, name, timezone, requireApproval. |
get_brand_context | The brand’s voice, audience, rules, pillars, hashtags and slots. Whatever the brand hasn’t filled in is left out and listed as missing: it is never guessed. | Read | brandId | brand, description, industry, website, voice, pillars, defaultHashtags, postingSlots, missing[]. |
list_accounts | The connected social accounts per brand. Their ids are used as accountIds when creating a draft. | Read | brandId (optional) | accounts[]: id, brandId, platform, handle, status. |
get_calendar | Posts between two dates (up to 92 days), read in an explicit time zone. | Read | from, to (AAAA-MM-DD); brandId, timezone (optional) | range, posts[] (status, scheduledAt, dateMeaning, localDateTime, caption, targets), note. |
list_posts | The most recent posts, filterable by brand and status. | Read | brandId, status (draft, in_review, changes_requested, approved, planned, scheduled, published, failed), limit (1–100) | posts[], note. |
get_analytics | Per-account totals over a range (up to 400 days): impressions, engagements, link clicks, video views and followers. Whatever a network doesn’t report is left out: it never shows as zero. | Read | from, to (AAAA-MM-DD); brandId (optional) | from, to, note, accounts[]. |
create_draft | Saves a draft with its text and, optionally, accounts and a proposed date. It never schedules it, sends it for approval or publishes it. | Create drafts | brandId, caption, idempotencyKey (required); accountIds, proposedAt, internalNote | post (id, status: "draft"), reviewUrl, note, replayed. |
Each block is an example request with placeholders between < > where your own data would go. Below it, the shape of the response: the fields the server returns, with no values.
What you’d tell a connected assistant and the tool it would call:
You: List my FLOW brands.
Tool: list_brands
Arguments: {}{ "brands": [ { "id": "<id>", "name": "<name>", "timezone": "<time zone>", "requireApproval": <true | false> } ] }The time zone is explicit; by default, the brand’s.
Tool: get_calendar
Arguments: { "brandId": "<id>", "from": "2026-10-01", "to": "2026-10-07" }{
"range": { "from", "to", "timezone", "startsAt", "endsBefore" },
"posts": [ { "id", "brandId", "status", "scheduledAt", "dateMeaning": "<proposed | scheduled | published>", "brandTimezone", "localDateTime", "caption", "targets": [ { "accountId", "platform", "format" } ], "updatedAt" } ],
"note"
}A new idempotency key per draft; the same one if you retry that draft, so two aren’t created.
Tool: create_draft
Arguments: { "brandId": "<id>", "caption": "<text>", "proposedAt": "2026-10-03T10:00:00-04:00", "idempotencyKey": "<a new uuid per draft>" }{
"post": { "id", "brandId", "status": "draft", "proposedAt", "brandTimezone", "proposedLocal", "targets": [ … ] },
"reviewUrl": "<link to review it in FLOW>",
"note": "Draft only: not scheduled, not submitted for approval, not published. A person reviews it in Social Flow.",
"replayed": <true | false>
}There are no packaged skills for agents yet. What exists are the seven tools above and example instructions for an agent (Grok Bot). This is how they map to each area, with their limit:
| Area | What it allows | Tools | Status and limit |
|---|---|---|---|
| Social | Read the calendar and save drafts for a brand’s networks. | get_calendar, list_posts, list_accounts, create_draft | Tools ready; not connectedPublishing and scheduling don’t exist over MCP, on purpose. |
| Content | Write in the brand’s voice and leave the text as a draft to review in FLOW. | get_brand_context, create_draft | Tools ready; not connectedThe app’s AI assistant is separate and is still off. |
| Design | Build carousels and pieces with the brand’s brand kit. | No MCP tool. | No MCP toolIt’s done inside the app (Super Flow AI). AI generation is built and off. |
| Branding | Look up each brand’s context and rules before creating. | list_brands, get_brand_context | Tools ready; not connectedRead only: editing the brand over MCP doesn’t exist. |
Example instructions for an agent: “Look up the accessible brands, pick the right one, read its context and calendar, prepare the requested content, save drafts when the user asked for it and return their real IDs and links. Don’t announce posts, approvals or metrics that FLOW hasn’t confirmed.”
A connected assistant acts as you, with your role and only the brands you authorize. These limits are covered by automated tests.
No. The connectors are implemented and tested locally, but not yet deployed or checked with a real assistant. When they are, you’ll see it here and in the product status.
No. It will only read your brands, calendar and metrics, and save drafts that you review in FLOW. Publishing, scheduling and approving stay in a person’s hands.
No. The calendar, approvals and publishing work without AI. The assistant and the connectors are optional.
There is no AI provider configured yet, so today it uses none. Any provider and model names that appear in FLOW are trademarks of their owners.
Guided access: we review your request, set up your first brand with you and send your invitation.