Building an AI social media manager on Second Studio
A lot of the people who ask me about Second Studio are not making videos themselves. They are building agents that make videos for someone else: a social media manager for small businesses, a marketing assistant inside a CRM, a content agent for a creator. This post is the architecture I would recommend for that kind of product.
Connect over MCP or REST
If your agent runs in an MCP client, one line connects it:
claude mcp add --transport http second-studio \
https://second.studio.hellofaro.io/mcp \
--header "Authorization: Bearer ss_live_..."
If you are building your own service, the same key works for the REST API. Every step is a tool: analyze_style, create_edit, record_url, write_script, create_explainer, create_post, get_job.
Split the work into a weekly loop
The loop I like is boring on purpose:
- Monday: the agent reads the business's calendar or product feed and proposes three post ideas.
- For each idea it picks a tool. Fresh footage from the owner goes through
create_editwith a saved style. A new feature or offer goes throughwrite_scriptthencreate_explainer. - It polls
get_jobuntil each render succeeds. Renders that fail quality checks fail the job and refund it, so the agent does not need its own video QA. - It drafts each post with
create_postfor a scheduled time. - A person approves or rejects the drafts. Rejected drafts refund their credit.
Keep a person on the publish button
Drafts are created with an agent key and approved with a separate admin key. I think that split matters for any product that posts on someone's behalf. Your agent can do all the work, and the business still decides what goes out under its name.
Budget per customer
Give each customer, or each seat, its own key with a spend cap. Caps are checked before a job starts, so a loop that goes wrong stops at the cap instead of running up a bill. A week of three edited posts is 3 edits (15) plus 3 drafts (3), so 18 credits, plus 2 once per new style.
Let the agent report problems
If something looks wrong, the agent can call report_issue itself, with the job id. That comes straight to us.
The agent builders page and the developer quickstart have the rest. If you are building one of these, tell us about it.