The agent creative playbook

Your agent writes the brief and then stops

The tool surface, grouped by job

This is the full set of tools an agent can call on your Advibly account over OAuth. Read it as a capability map: if a job is not on this list, the agent cannot do it, and an instruction that assumes it will fail politely instead of loudly. Generation tools spend credits against your balance. Tools marked free spend none.

From zero to the first returned asset

There are three ways to connect an agent, and they are not competing options. They suit different moments. Do steps 1 and 2 once, then pick whichever connection matches where you actually work.

1. 1

   1. Create the account before you touch any agent config

   Sign up at advibly.com. New accounts get 2 free starter credits, which is enough to run a real generation and look at the output before paying for anything. Access for agents is OAuth. There is no API key to mint, copy or leak, which also means there is nothing useful to paste into a config file yet.

2. 2

   2. Build the brand context first, because everything reads from it

   Paste your website, App Store listing, Play listing or storefront URL into onboarding. Advibly scans it for identity, palette, fonts, logo and imagery, renders a brand kit sheet, pulls the product catalog for store brands, and starts the deeper research that produces the brief and dossier. You can do this without opening the app at all: an agent that is already connected can call advibly_onboard_brand with the same URL and run the identical onboarding in chat. Skipping this step is the single most common reason agent-made creative comes back looking generic.

3. 3

   3. Hosted MCP: the default, and the right one for a chat client

   Add the hosted endpoint at advibly.com/mcp to your MCP client. The client opens a browser, you sign in, you approve, and the tools appear. No key, no secret, no server to run. Per-client setup guides live at /mcp-setup for Claude, Claude Code, ChatGPT, Cursor, Codex, VS Code and other MCP clients. Choose this when the agent you work in is a chat or IDE client and you want the tools available in every conversation.

4. 4

   4. The CLI: the right one for a shell, a script or a pipeline

   Run npx advibly login. The CLI is an MCP client too, so it goes through the same OAuth chain and inherits the same tools: every tool becomes a subcommand, so advibly_generate_image becomes advibly generate-image --prompt "..." --brand-id "...". Two commands are CLI-native because a browser cannot do them well: advibly upload takes a local file through a presigned PUT, and advibly download writes result URLs to disk. Pending generations are followed until they finish unless you pass --no-follow. Choose this when you want files on your filesystem, when you are scripting a batch, or when there is no chat window in the loop.

5. 5

   5. WebMCP: the right one when the human is already in the app

   WebMCP is the in-page surface. Instead of an agent calling your account from somewhere else, it drives the tab you are already looking at, on the session and the live data the page already holds. Tool names are unprefixed there, so it is generate_image rather than advibly_generate_image, and every call moves interface a human can see. Chrome exposes the API behind chrome://flags/#enable-webmcp-testing and ChatGPT's in-app browser has it built in; elsewhere a polyfill installs it. Anything that spends credits stops on an approval card first, and declining is a normal answer rather than an error. The live inspector at /webmcp reads the real tool surface of whatever tab it is open in. Choose this when you want the agent to operate the app in front of you rather than work in the background.

6. 6

   6. Install a skill so the agent stops improvising the pipeline

   Run npx skills add advibly/skills. Each skill packages one complete production method: concept, script, storyboard, model choice, generation order, voice, assembly and the review gates in between. That matters because the hard part of a video ad is not any single tool call, it is the order and the failure handling. Skills are free and open source. Running one spends normal credits like any other generation. Browse them at /skills.

7. 7

   7. Prove the loop with one cheap generation

   Ask the agent: list my brands, check my credit balance, then generate one 1:1 image for brand X of the product on a plain background, and give me the URL. That single instruction exercises brand resolution, the credits read, a generation and the result fetch. If the URL comes back and opens, everything downstream in this playbook will work.

8. 8

   8. Write the reporting rule into your instructions once

   Add a standing line to your project instructions or system prompt: always report the credit estimate and my balance before generating, and always finish by giving me every result URL in one list. An agent that has not been told to report a URL will cheerfully announce success without ever fetching one.