How AI agents use scatterpost

scatterpost is built for an agent to use directly, not only through a dashboard. This page explains what a new model sees the moment it connects, so you can tell whether your own agent is using it well.

What the agent sees on connect

An agent learns how to use a tool from five places, and scatterpost fills every one of them:

Server instructions. Returned the moment the connection opens: what scatterpost is, the canonical-first rule, a short recipe for every standard flow, and what never to do.
Tool names, descriptions and input schemas. Every one of the 18 tools states what it does, when to use it, when not to, and what it returns, with a description and an example on every parameter that needs one.
MCP prompts. Four ready-made starting points for the main workflows: publish_article, schedule_week, connect_first_channel and weekly_results.
MCP resources. scatterpost://guide/getting-started (the full onboarding guide, in Markdown) and scatterpost://reference/platform-limits (character limits, link handling and thread support for each platform).
Error messages. A failed tool call never just reports what went wrong: it names the next step, for example which field to fix, which tool lists valid channel ids, or which plan to upgrade to.

The standard recipes

The same flows appear in the server instructions, the getting-started resource and the prompts, so an agent sees them however it looks:

First run: list channels; if none, connect one in the dashboard, or have the agent connect an existing site itself with site_url, or deploy one of the four ready-made templates and connect it, before drafting anything.
Draft, preview and publish now, once the user says go.
Schedule for later, in the workspace's own timezone.
A short social post versus a full blog article.
Bulk scheduling across channels or articles (Team plan).
Approvals: only a workspace owner releases a held publication.
Reading results and analytics after the fact.

Setting up a blog without the dashboard

An agent can connect a website channel on its own: scatterpost_channels with action "connect", platform "website" and site_url derives the receiving endpoint itself, and action "test" checks the credentials before anything is drafted. For a founder with no blog yet, the scatter-blog-setup skill in scatterpost's Claude Code plugin walks the agent through picking one of the four ready-made templates, deploying it to the founder's own Vercel account, generating the webhook secret locally and connecting it, always asking before anything deploys and never printing the secret in chat.

An example conversation

User: Publish the retro I just wrote to my blog and cross-post it to Dev.to.

Agent: [calls scatterpost_channels, action "list"]
       Found two channels: "Main site" (website) and "Dev.to".
       [calls scatterpost_draft with the retro's title and body]
       Saved as a draft. Want me to publish it to both channels now?

User: Yes, go ahead.

Agent: [calls scatterpost_publish_now with both connection ids]
       Your website published at https://example.com/blog/sprint-12-retro.
       Dev.to published at https://dev.to/you/sprint-12-retro, with the
       canonical URL pointing back at your site.

Connecting an agent

See the MCP server docs for the connection details: the remote endpoint, the stdio command for Claude Code, API keys and scopes, and the full tool list.