What this guide does: Take an app from setup to a reviewed post and confirmed delivery using only your chat app and a browser. No CLI or installed skills needed.
You tell the agent what you want; it handles the tool calls below. The tool names are here so the agent knows the order. Setup should only ask for the connection the current task needs.
Connect Sprid
Choose your chat at https://sprid.studio/start#chat and follow its steps. Your chat may call this an app, connector or integration. The address is https://api.sprid.studio/api/mcp?surface=core. Sign in to Sprid and authorize the right workspace.
Agent: when available, start with get_marketing_plan and reuse its prepared work and next step rather than starting over. Discovery can start from a website or store URL, with no social account.
Pick the app and account
Agent: call get_capabilities, then list_apps and list_accounts, and reuse what exists. If the app is missing, get its description or website/store URL, confirm it and create it with upsert_app_profile. A content account owns voice and media; a channel is one exact social destination. Never pick a handle just because it is first in a list. A missing account is created in the dashboard.
Connect a destination or data source
- Social channel:
get_documentationwith the platform name, thenconnect_channel. The person opens the returned link and approves in the browser;channel_statusconfirms the saved result. - Store or analytics: the person enters identifiers and the key in the secure form at
https://app.sprid.studio/settings/app-profiles. Credentials never go in chat, tool arguments or reports. A saved-key indicator only proves configuration;get_marketing_reviewdoes a live read.
Bring images or a finished video
Draft copy in the chat, in the account’s voice. No extra model key needed.
If the chat exposes the attached files, pass them to import_assets: each file has download_url and file_id, optionally mime_type and file_name. Public HTTPS URLs and existing {kind, id} library references also work. Limits: PNG, JPEG, WebP or finished MP4, up to 32 MiB each. Each position reports on its own; retry the missing ones before building the post. Files are copied to the account and deduplicated by bytes; temporary download URLs are not kept.
A sandbox: URL or local path can’t be fetched. For an image the chat generated, use the existing file if the host exposes it (not verified for every host); otherwise download it and use the browser upload. Never ask the agent to recreate an image or transcribe binary data.
Fallback: browser upload. Call create_upload_session with the account and a new UUID requestId. The person opens the link and picks the files. Then read get_studio_job with the returned ID; its result keys are zero-based positions. The same session and asset IDs work from any other chat. No image-generation charge.
Create and review the post
Call create_post_from_assets with a new UUID requestId, the account, ordered {kind, id} assets, captions and aspect ratio. Reuse the same ID and inputs if a response is lost.
presentation: "finished"keeps artwork as-is: no text overlay, no added music. Images fit inside the canvas, which can add margins.presentation: "editable"allowstextandsubtitleon image slides.- One finished video becomes one reel. Never mix a video with image slides.
Edit with update_post (captions) and update_slide or batch_update_slides (slides). Call preview_post and check every slide or the video: captions, order, crop, legibility. Preview links work even if the chat can’t embed them. A preview is not approval to publish.
Make a reel from photos or text
Check get_capabilities for hosted creation first. create_reel takes ordered scenes (imageId, text, durationMs), up to 120 seconds, and makes a silent MP4. It costs 3 cents per started render minute, or uses the plan allowance: check spend_status, tell the person, and get their go-ahead before sending confirmed: true. Save the job ID and read get_studio_job. Retrying with the same ID never starts a second render; a stopped job reports needs_attention instead of charging again.
App captures, simulator recordings and your own build scripts run outside remote MCP. Upload their finished files.
Approve and confirm delivery
The person opens the returned /p/<postId> link, reviews the real media and captions, and picks the channels, time and platform options. TikTok privacy and interaction choices have no defaults and are made on that screen. An agent-supplied confirmation flag is not a human click.
After approval, read get_publication_status. Keep draft, scheduled, failed and delivered apart, and report delivery only with the platform receipt. Read a failure before retrying. Never repeat a publish call just because a response was lost.
For results, read get_documentation for marketing-review. Keep the same app, account and post IDs across clients. next_actions gives the next step; an empty verb means follow its browser link.
Limits and troubleshooting
- Attachment can’t be read: use the browser upload, not the CLI.
- Missing account or service: set it up in the dashboard. Never paste credentials into chat.
- Hosted creation: check
get_capabilitiesbefore promising it. Store screenshots and store releases are local unless it reports otherwise. - File transfer: whether a host exposes uploaded or generated files has to be checked per host. A supported schema doesn’t prove it.