Publish from an AI agent

Why Dravo suits agents

An agent does best with one predictable contract and machine readable failures. Dravo gives both: a single publish call for every network, and errors that name what to fix. There is also a raw Markdown mirror of every docs page (the View .md button at the top of each page), which is easy for an LLM to ingest.

The minimal loop

Give the agent exactly one capability: publish text (and optional media) to a set of accounts.

async function publish({ accountIds, text, mediaUrls = [] }) {
  const res = await fetch("https://api.dravo.dev/v1/publish", {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.DRAVO_API_KEY}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({ account_ids: accountIds, text, media_urls: mediaUrls }),
  });
  if (!res.ok) {
    const err = await res.json();
    throw new Error(err.detail); // actionable message the agent can act on
  }
  return res.json(); // { job_id, status }
}

Reading errors and retrying

When a request is rejected, the body contains a detail string describing the problem (for example a missing field or text over the limit). An agent should parse it, adjust the input, and retry. When a job is accepted but a single account fails, the per account result carries a stable error_code plus a human message, so the agent can decide whether to retry that account or skip it. See handling partial failures.

As an MCP tool

Wrap the publish call as a single MCP tool named something like publish_post with inputs account_ids, text and media_urls. Keep the surface tiny: one tool, clear inputs, and let Dravo's response drive the agent's next step. Point the model at the raw docs (/docs/quickstart.md) so it knows the contract.