> ## Documentation Index > Fetch the complete documentation index at: https://docs.autogrowth.farm/llms.txt > Use this file to discover all available pages before exploring further. > ## Agent Instructions > AutoGrowth runs social accounts on real phones. Everything a page describes happens on a physical device in the customer's own fleet, never through a social platform's API. > The Companion is a macOS desktop app that runs on the customer's own Mac and manages the phones over USB — iPhones through WebDriverAgent, GrapheneOS phones through ADB and scrcpy. Nothing runs while the Companion is closed. > The Companion builds, signs and installs WebDriverAgent on each iPhone itself. Never tell a customer to sign WebDriverAgent by hand in Xcode; Xcode is a prerequisite the Companion drives, not a step the customer performs. > Autogrowth MAX is a one-time, invite-only lifetime license bought from the organization's credit balance, not a subscription tier. It is required to run any automation, workflow or Incubator program. Without a price set for that organization, the purchase is refused. > An automation is one of AutoGrowth's built-in definitions; a script is a graph the customer builds in the Script Builder. The Script Builder is locked platform-wide, so never present building a script as something a customer can do today. > An API key is a bare 64-character hex string with no prefix. It is shown once, at creation, is sent as `Authorization: Bearer `, and can never be read back. > An API key sees only the key owner's own accounts, content, collections, prompts, scripts and executions. Devices and tags are organization-scoped, and a non-owner sees only the devices assigned to them or carrying one of their tags. The automation catalog is global. > There is no pagination. List endpoints return the whole set; the only limiting parameter anywhere is `limit` on `GET /v1/executions`, which defaults to 50 and has no offset to page with. > Remote control and the live screen are dashboard-only, over WebSocket and MJPEG. Neither is part of the public API. > The MCP server is hosted at https://mcp.autogrowth.farm and authenticates with OAuth. It is not an npm package and there is nothing to install locally. Its tools inherit exactly the scope of the API key behind them. > Android support is partial and GrapheneOS-only: those phones appear in the fleet and can be watched and remote-controlled, but they cannot run scripts or automations. > Never document Hi Katie or any persona feature. When a page and the code disagree, the code wins — check a claim against the source before repeating it. # MCP limitations > What the MCP server will not do, and the scoping rules that decide what an assistant can reach. The MCP tool set is a slice of the product, not all of it. Everything below is a fact about the current server, not a promise about a future one. ## Scheduling is one-way An assistant can create runs. It cannot undo them. * **No cancelling.** There is no tool that cancels a pending or running execution. * **No deleting.** A run scheduled by mistake stays until it runs, and has to be removed from the dashboard's **Schedule** screen. * **No editing.** A run's time, device or inputs cannot be changed after the fact. Delete it in the dashboard and schedule a new one. Cancelling and deleting do exist on the REST API — see [Cancel execution](/api-reference/executions/cancel-execution) and [Delete execution](/api-reference/executions/delete-execution). Editing a scheduled run exists nowhere. ## Results are shallow `list_schedules` reports a run's `status` and, when it failed, the error recorded on it. It does not reach the artifacts. * **No logs.** The step-by-step log, the error screenshot and the screen dump a failed run uploads are not exposed to any tool. An assistant cannot tell you why a run failed beyond the one-line error. * **No history beyond executions.** Account follower history, and the dashboard's aggregate stats, have no tool. ## Whole areas have no tools | Area | Status over MCP | | -------------- | ---------------------------------------------------------------------------------------------- | | Collections | No tools at all. Uploads cannot target a collection, and content cannot be moved between them. | | Prompts | No tools at all. | | Accounts | Read-only. No pausing, resuming, or reading an account's scrape history. | | Content | Upload and list only. No starring and no deleting. | | Tags | Readable only as part of a device. No tag list, and no assigning tags or teammates to a phone. | | Storage | No way to check how much of your allowance is left before an upload fails. | | Remote control | Not on MCP, and not on the API either. The live screen and remote control are dashboard-only. | ## Results can be very large There is no pagination anywhere. `list_devices`, `list_accounts` and `list_media` return the full set every time, and `list_schedules` caps at `limit` with no offset to page through the rest. On a fleet of a few dozen phones that is a lot of JSON, and it lands in the assistant's context window in one piece. Narrow the request — filter `list_schedules` by device and date range, filter `list_accounts` by platform — rather than asking for everything and letting the assistant sort it out. `list_media` also strips file and thumbnail URLs, so an assistant can tell you what is in your library but cannot see it or hand you a link to it. ## Scope is the key owner's, not the organization's A connection acts as the person who approved it. Accounts, content, scripts and executions are limited to **your own** rows, even when you own the organization. Devices and tags are organization-wide but filtered by what you can reach. Full rules in [What a key can see](/api/authentication#what-a-key-can-see). ## The wider API The REST API covers most of what the tools do not: cancelling and deleting runs, collections, prompts, account pause and history, content starring and deletion, storage usage, and device tag and user assignment. Base URL, keys, and a three-step start. Every endpoint, field by field.