Docsbook documentation

Give your agent a goal. Get back an opportunity, a number, and a verdict.

Say where you want the business to get to. Docsbook researches what is actually there, sizes the opportunity from evidence, and turns it into experiments that each carry a number they are expected to hit — then reads, on the date, whether they hit it. It runs inside the agent you already use, and everything it writes lands in your own repository as a commit you review.

Quick start How it works MCP server Pricing

Paste one prompt into Claude Code, Cursor, Codex or any client that speaks MCP. It connects the server itself, writes a Docsbook rule into its own memory file, and takes a first project live.

What people bring it

Be found

"More qualified organic traffic." Read as clicks and impressions per intent.

Be recommended

"Be named when assistants answer my category." Read as citations across a fixed question set.

Turn readers into customers

"Make more readers reach the product." Read as goal completions per page.

Open a new market

"Reach people in another language." Read as traffic and rank per locale.

The loop, in eight stages#

The stage most tools skip is the fifth: writing down what should change, and by when, before doing the work. Without it the seventh stage has nothing to compare against.

Goal

One sentence from you — the outcome you want, not a brief or a backlog.

Research

Six classes of evidence: what your product does, the demand around it, who ranks for it, what assistants answer, what your readers searched for and did not find.

Opportunity

One row per thing people look for: how many of them, what your docs show them today, who gets the click instead, and what winning it is worth.

Hypothesis

Competing ways to take the same opportunity, each argued from the readings rather than from a checklist.

Expected effect

The number it is expected to hit and the date it will be read on — written down before any page exists.

Experiment

The change is written into your repository as a commit or a pull request you review like any other.

Actual result

On the date, the reading is taken and compared with the number written beforehand. Confirmed or rejected.

Learning

A rejection counts: what it eliminated is kept, so the next bet does not repeat it.

The experiments run on documentation because it is the surface an agent can change without asking anyone for access, and the one whose effect is readable from outside — search indexes it, assistants quote it, buyers decide on it, and every page reports its own traffic, referrals and failed searches.

The surface they run on#

Four stages, and an experiment can be read at any of them. Pick the one you are asking about.

1Publish

Source to a public URL, in about five minutes.
2Get found

What ships for search engines without configuring anything.
- SEO
- llms.txt
3Get quoted

What makes a passage an assistant can cite.
4Measure

Which pages were read, and where readers stopped.

Start in one conversation#

There is no signup flow in front of it. Point the agent you already work in at one endpoint and say where you want to go.

claude mcp add --transport http docsbook https://docsbook.io/api/mcp/server

Codex, Cursor, Windsurf, Cline and any other MCP client use the same endpoint — see MCP server for each. Would rather see the surface first? Publish a site in about five minutes.

Start here

Overview

What Docsbook does, and what it costs

Quick start

Source to public URL

Concepts

Workspace, balance, indexing

Use cases

The jobs teams hire docs for

How we prove it

The evidence rule every page follows

Run the loop from your agent