Docs

Sitearc is two things in one account. It is the research stack you open Semrush or Ahrefs for. It is also a loop. That loop turns research into a change, gets your approval, and ships it as a pull request. Then it measures what shipped.

Both halves are reachable three ways. You can work in the app. You can point an agent at the same data over MCP. You can install the Claude Code plugin and get the tools plus eleven guided skills.

What to read first, by role

Solo developer running several sites

Start with Getting started. Add one site, let the crawl finish, and read the first report. Connect Search Console next, because nothing gets measured without a measurement source.

Then read MCP tools and connect your agent. If you use Claude Code, read the plugin doc instead and install the plugin. The seo-loop skill is the one that runs the whole cycle.

Finish with Credits and plans. Research calls cost money upstream, and the doc says what spends and what does not.

Agency running client sites

Start with Connect a site. It lists every connection, what each one reads, and what breaks when a client cannot give you access. Client sites are usually missing one source, so read the missing-source sections carefully.

Then read Credits and plans for the per-site economics. Read the agencies page for roles, per-site history and what is planned.

If your clients are on WordPress, read the WordPress plugin doc. That connector is being built, and the page describes the plan rather than a shipped feature.

Marketing team that does not own the repository

Start with Getting started, then work the app. Every proposal carries the exact change and instructions to apply it by hand. Completing it that way is recorded as applied manually.

Read the tier ladder in MCP tools before you quote a number to anyone. A measured number and an estimated number do not carry the same weight. Sitearc labels which is which.

The seven docs

Doc	What it covers
Getting started	Sign up, add a site, first report, first proposal
Connect a site	Every connection, what it reads, what it needs
MCP tools	The full tool table, OAuth, rate limits, refusals
Claude Code plugin	Install, the eleven skills, the honesty rules
WordPress plugin	The connector being built, and its endpoints
Credits and plans	Credits, budget checks and current beta capacities
Credits	The open source project Sitearc is built on

The rules that never bend

Four of them are worth knowing before you read anything else.

Nothing ships without a person. There is no approval tool for an agent to call. The module that records a decision cannot be imported from the agent-facing code. A test fails if that changes.

Every value carries a tier. Tier one is measured from Search Console or Analytics. Tier four is inferred, which means a hypothesis. Nothing renders without a tier.

Evidence is required, not encouraged. A proposal with no evidence is refused by the tool. It is not scolded by a prompt and then allowed through.

A verdict is computed, never stated. conclude_experiment applies the standard and returns the result. There is no field where an agent writes its own outcome.

What to do when a result is inconclusive

An experiment that finishes without a clear effect is marked inconclusive. That is a real result, not an error.

Most changes land there, and a tool that cannot say “this did nothing” will tell you everything worked. Sitearc records the inconclusive verdict and moves to the next proposal. A learning citing it is stored as a hypothesis.

Do not go looking for the directional read the verdict withheld. A band, a gap or a trend after that word is the answer the standard declined to give.

What is being built

Scheduled refresh, the recommendation queue, the feedback tool, the content engine and the chat are being built. So are paid checkout and credit pack purchases. The WordPress plugin and GitHub App are also being built.

Each doc labels the parts that are not shipped.