How an AI agent can run a book-publishing workflow (MCP, CLI or API)

Most SaaS products assume a human at a browser: they gate signup behind CAPTCHAs, hide functionality in dashboards, and answer scripts with challenge pages. An agent hitting those walls stalls silently. Agent-operable means the opposite — every step the workflow needs is reachable through a documented, machine-friendly interface that returns structured data, not HTML.

For DraftToDone that means a session token an agent can hold, JSON responses with clear error codes, and generation that runs server-side so the agent can fire a job and poll for it later instead of holding a connection open for 40 minutes. The agent never has to render a page or solve a puzzle.

The design rule is simple: anything a paying customer can do from the dashboard, an agent can do from the API. The single deliberate exception is payment, for reasons covered below.

The MCP server at app.drafttodone.io/mcp is the native choice for frameworks that speak Model Context Protocol. Hermes Agent discovers its tools at startup; Claude clients add it with one command; any MCP host registers it as a remote streamable-HTTP server. The agent then calls named tools — sign_up, get_checkout_url, create_book, list_books, get_download_links — like any other tool in its registry.

The CLI (npx drafttodone, or a single downloadable file) suits agents that already have shell access and prefer to run commands. Every command accepts a --json flag so the output is machine-parseable. It is a thin client over the same backend, so behaviour is identical to MCP.

The REST API is the lowest common denominator: plain HTTPS with a cookie jar, usable from any language or a bare curl. Agents without MCP support or a shell fall back to this. All three surfaces enforce the same rules, rate limits and credit logic — there is no 'privileged' path.

Step one, account: the agent calls sign_up with the user's email and a password, and receives a session token it stores. Step two, funding: it requests a checkout URL and hands that URL to the human — this is the one blocking step. Once payment clears, credits appear automatically.

Step three, generation: the agent calls create_book with a specific niche and language. Specific beats broad every time — 'beginner strength training for women over 50' outperforms 'fitness.' The job runs server-side for roughly 30 to 45 minutes. Step four, polling: the agent checks list_books every minute or so until the status reads completed and the progress counter reaches four of four.

Step five, delivery: the agent fetches short-lived download links for the four files and saves them for the user. Step six, the human's turn: the user reviews the content and publishes from their own Amazon KDP account. An optional step seven — after the user posts about the book on X — lets the agent claim a free weekly credit.

Payment is deliberately not automatable. The agent generates a Stripe Checkout URL, but a person must open it and complete the transaction. This is not a missing feature; it is a guardrail. An autonomous agent with unattended spending authority is a liability for its owner, and DraftToDone refuses to be the service that drains a card while nobody is watching.

Publishing is the second human checkpoint, this time by Amazon's rules rather than ours. The user uploads the files to their own KDP account, sets price and categories, and — importantly — declares the AI-assisted content, which KDP requires. The agent produces the assets; the human owns the listing and the compliance.

Framed correctly, these checkpoints are what make the workflow safe to run on a long leash. The agent does the tedious production work continuously; the human keeps authority over money and publication. That division is the point, not a limitation to engineer around.

DraftToDone produces files; it does not produce guaranteed income. No tool can promise Amazon approval, search rankings or sales, and any service that does is lying. The realistic value is leverage: turning a niche into a complete, consistent book package in under an hour, repeatable across a catalog, at a per-book cost of roughly one to two and a half euros on subscription.

Quality still needs a human's eye. Generated manuscripts are a strong first draft, not a finished product exempt from review — reading before publishing is both good practice and, for AI content, an ethical baseline. Agents should present the output as a draft to approve, never as guaranteed-sellable on arrival.

And the market is real: a flood of low-effort AI books has made buyers and platforms warier, so the books that earn are the ones with a genuine niche, honest quality and a real cover — exactly the parts a human still steers. An agent that treats DraftToDone as a volume spam engine will lose; one that treats it as a production accelerator behind human judgment has a workable system.