Get started with your AI agent

You open your AI agent in an empty folder and paste one prompt. The agent creates the project, checks the machine, runs your brand interview, builds the first video and prepares publishing. You only approve and do what only you can do: create the account, store keys, connect the channel and accept the license. This page is the route; the details of each piece are on the pages it links to.

0. Who it is for and what the agent needs

On both paths, the remote praeko MCP server publishes (praeko.app/mcp), always in two steps: the preview and the confirm. They are two servers with different roles: the local one (praeko-studio) creates and renders the video; the remote one (praeko) looks at your account, uploads the media and publishes. Connect your agent to the remote one from its page in AI agents.

On the machine: Node 22, ffmpeg and Chrome (praeko studio doctor checks and says how to fix; see the Studio requirements). The agent, the voice and the render are on you; Praeko neither bills nor brokers the agent.

1. The starting prompt

Create an empty folder, open the agent in it and paste the prompt below, replacing what is between < and > with your answers. The agent creates the project with init --yes (no questions in the terminal), runs doctor, runs the brand interview by talking and never asks for or reads keys.

You will take me from zero to the first video of my channel with Praeko Studio, in this empty folder. Do everything you can yourself. Rules:
- Never ask for, read or print keys (voice, Praeko or any network). If a key is needed, tell me the command to run in MY terminal.
- Do not spend voice credit or publish anything without my yes, after showing me what will happen.
- If you have no terminal, stop and tell me: the path without a terminal uses the local Studio MCP server.

1. Ask which AI agents I use (it can be more than one, for example claude,codex) and pass the list in --agent. Then create the project (use my answers below; leave out what I do not know):
   npx @praeko/studio@latest init <folder> --yes --agent <agents-i-use> --name "<channel name>" --handle "@<handle>" --lang en --about "<what the channel is about>" --audience "<who it is for>" --tone "<tone of voice>" --accent "#f4b942" --background "#14110f" --font "Arial Black"
2. Go into the folder and run `praeko studio doctor` (if the `praeko` command does not exist, use `npx praeko-studio doctor`). Explain anything missing (Node 22, ffmpeg, Chrome, fonts) and how to fix it.
3. Read the project instructions. If `brief.md` does not exist, or `playbook.md` still has TODO, run the brand interview BY TALKING: one short question at a time, all skippable (name, audience, tone, colors, font, up to 3 references, voice: only the nickname and the provider). Write `brief.md` and `playbook.md` and adjust `brand.json`. Tell me what you filled in.
4. Stop and hand me the list of what ONLY I can do (Praeko account and key, voice key, channel, license notice), with the exact commands. After my "done", go on to the first video.

Each answer becomes an init flag: --name, --handle, --lang, --about, --audience, --tone, --accent, --background, --font. If you already have a brand folder with tokens, use --brand <folder>. Whatever is left out, the agent asks in the conversation and writes to brief.md, playbook.md and brand.json. Studio's instructions for your agent (the *skill*) are written by --agent; see Instructions for your agent.

2. What only you do

The agent prepares the list and the commands; you run them in your own terminal (not in the agent's chat: some clients let you run a command with !, but check yours). No key goes through the conversation.

Praeko account and key

• Sign in at praeko.app with Google and create a key under Developers, API keys. The recommended scope is write: it uploads media and creates posts, without touching keys or webhooks. See Authentication and scopes. Praeko is a closed beta today: the waitlist gives access.
• Store the key without showing it to the agent: keep it in the PRAEKO_KEY variable of your terminal (from a password manager, not written in a project file) and run echo $PRAEKO_KEY | praeko studio schedule --via praeko --key, which stores it in the macOS Keychain or the secrets file. If you use the remote MCP server through sign-in (OAuth) you do not need it to publish.

Voice key (only for the final voice)

The first video uses a draft voice and needs no key. For the final voice, use your own voice (or one you are authorized to use) on your own ElevenLabs account. In your terminal:

praeko studio voice account add principal --voice my-voice --voice-id <YOUR_VOICE_ID>
praeko studio voice balance

Another way is echo $ELEVENLABS_API_KEY | praeko studio voice config key elevenlabs/principal. See Keys and costs.

Channel and license

• Connect the channel (YouTube, Instagram or TikTok) under Channels in the panel: signing in to the networks is yours, the agent does not do it. All three networks still depend on the apps' approvals (Meta and TikTok): if a connection does not show up or TikTok only publishes as a draft, that is today's limit. YouTube goes up private and scheduled.
• Accept the license notice: praeko studio licenses --accept records that you read it (praeko studio docs licenses shows the text).

3. The first video

With the project ready and the list above done, ask for a test Short. The voice is a draft (estimated and silent, free, no network), so nothing is spent; check warns that it is not for real publishing, and that is expected.

Make a test Short of about 30 seconds about <TOPIC> in this Praeko Studio project. Read the project instructions, brief.md and playbook.md, and run `praeko studio docs episode-yaml` before writing the YAML.
1. Create the episode: `praeko studio new my-test --title "<title>"`. Write the script (script.txt) and show it to me. Go on only after my yes.
2. Draft silent voice, which costs nothing: `praeko studio voice my-test --quality draft`. Do not use `--quality final`: it spends characters from my voice account and only with my yes.
3. Write the scenes in episode.yaml and run, in this order: `praeko studio compile my-test`, `praeko studio check my-test`, `praeko studio render my-test --scale 0.5`, `praeko studio contact-sheet my-test` (look at the image and fix what is wrong), `praeko studio cover my-test` and `praeko studio package my-test`.
4. Show me the package to approve: the script, the contact sheet, the cover, the texts.md of each network and the check warnings (the draft voice is expected). Publish nothing.

What you approve here: the script (before any spend), the frame mosaic (contact-sheet), the cover and the texts.md of each network. For the final voice, the agent shows the text and the balance (praeko studio voice balance) and waits for your yes before --quality final. The AI content disclosure lives in publication.json (see Compliance).

4. Publish, always after your yes

The remote praeko MCP server publishes, in two steps: the first call returns only the preview and a confirmation_code (valid for 10 minutes and one use); the agent shows the whole preview and only repeats the call with confirm after your yes. If any argument changes, the preview is redone. Details in MCP server.

Publish the video my-test through the `praeko` MCP server, only after my yes at each step.
1. Call list_channels and tell me which channels you plan to publish to.
2. Upload the video and the cover: with a terminal, create_media_upload, run the curl it returns (the key comes from $PRAEKO_KEY in the environment; never ask for it or print it) and complete_media_upload. Without a terminal, studio_upload on the local Studio server.
3. Call create_post (or schedule_post, if I give you a time) WITHOUT confirm. Show the whole preview: channels, time in BRT and UTC, media, texts, warnings.
4. Only repeat the call with confirm when I say "go ahead and publish". If I change anything, ask for a new preview.

From the command line, without the remote MCP, praeko studio schedule <slug> --via praeko --when "dd/mm hh:mm" only simulates and --send really schedules (the agent never uses --send unless you ask). praeko studio post is direct publishing with your own credentials for each network, for advanced use. The video needs the final voice before it goes live.

5. Learn from the numbers

After a few posts, the agent reads the insights through the remote MCP server and proposes the next script. Read only: get_insights (views, likes, comments, followers, with comparison), get_retention (retention, impressions and CTR per post), list_posts and get_calendar. A missing number comes back as null, never 0. Reading and replying to comments is not a server tool: it stays in the panel and the API.

Read how my channel is doing through the `praeko` MCP server and propose the next script. Read only, create nothing.
1. get_insights for the last 28 days, compared with the previous period, and get_retention for the same posts.
2. list_posts with state "sent" to tie each number to a title.
3. Tell me in three lines what worked and what did not, quoting the numbers. Where a number comes back empty (null), say there is no data; do not guess.
4. Propose ONE script for the next Short, with the hook and the reason, and wait for my yes before creating the episode.

6. So you do not get lost

What the agent does and what you do

Common errors

More cases in Studio troubleshooting and Studio MCP.

Final checklist

1. Node 22, ffmpeg and Chrome ok in praeko studio doctor.
2. License read and accepted (praeko studio licenses --accept).
3. Praeko write key stored in your terminal, and the agent connected to the remote praeko MCP server.
4. Channel connected in the panel (with the caveat on Meta and TikTok approvals).
5. brief.md, playbook.md and brand.json reviewed by you.
6. Script, mosaic, cover and texts approved; final voice only with your yes.
7. AI content label checked (ai.synthetic) and set in the network's app when needed.
8. Post preview read in full before the confirm.

Links