Sitearc
Start free
Agents

How your agent works with Sitearc.

Connect the agent you already use. This guide covers how it signs in, what it can read and change, and where a person decides.

Start freeConnect your agent

130 tools

Only read
61
May change or spend
69
Can approve
0
Every tool in the MCP reference

Connect your agent.

One address, four ways in. Sitearc runs no AI of its own, so you bring the agent and pay no model markup.

Claude on the web

OAuth
  1. 1In Claude, add a custom connector.
  2. 2Paste https://app.sitearc.ai/mcp as its URL.
  3. 3Choose Sign in now, then Register automatically if asked.
  4. 4Sign in to Sitearc and allow access. Start a new conversation with Sitearc turned on.

After a Sitearc update, refresh the tools list in Claude's connector settings. Then start a new conversation.

Claude Code

OAuth

Add the server once, for every project.

$ claude mcp add --transport http sitearc --scope user https://app.sitearc.ai/mcp

Then run /mcp inside Claude Code to sign in.

Run claude mcp list first. If Sitearc is already there, keep using it.

Codex

OAuth

Add the server from your terminal.

$ codex mcp add sitearc --url https://app.sitearc.ai/mcp

Approve the Sitearc sign-in when Codex asks. Any other MCP client uses the same address.

An API key

Scripts and CI

For a client that can't sign in. Create a key in Sitearc's settings. Keys start with oseo_.

x-api-key: oseo_…

Or send it as Authorization: Bearer. Each key is read or write, and write is the default.

Sign-in
OAuth with the mcp and offline_access scopes. Access tokens last 24 hours. Refresh tokens last 30 days.
Reach
One connection covers every site you can open, across the workspaces in your account. Your role in each workspace applies on every call.
Read-only keys
A read key sees only the 61 tools that read. It cannot write, spend credits or publish, whatever your role.
Turning it off
Disconnect a client or revoke a key on the Agent access page. Other clients keep working.

What your agent can and can't do.

It reads the data you see in Sitearc. It can open pull requests, and a person merges them.

Reads, 61 tools

Search Console and GA4 reports, audits, rankings, saved research, forms, feedback, the brand kit and project memory.

Reading saved data never spends credits.

Changes or spends, 69 tools

Files ideas and proposals, opens pull requests, creates forms and reviews, sets up rank tracking and buys research.

Every tool that buys outside data counts here, even when it only returns data.

Never

  • Approve a proposal, a preview or a setup plan.
  • Merge or deploy a pull request.
  • Choose the result of its own experiment.
  • Resolve a confounder it logged.

There is no approve tool. A test in the app fails if one is added.

The tools, grouped by job

Where a person decides.

A change to your site passes three steps in Work. Proposals and previews are approved in the dashboard, signed in.

  1. 1

    Idea

    A site check, a comment, a form response or search data becomes an idea in Work. You accept it, or your agent can.

    Accepting only creates a proposal. Nothing changes yet.

  2. 2

    Proposal

    Your agent records the exact change, its evidence, and the direction and window it predicts. Only a person can approve it.

    Sitearc refuses to ship a proposal whose approval has no person recorded.

  3. 3

    Preview

    Your agent opens a GitHub pull request and submits it for review. You approve that exact revision.

    Sitearc fingerprints the pull request, its head commit and its preview link. If a newer revision is submitted before you decide, your approval is refused.

Demo data. Harbor & Pine is a fictional boatyard.

Shipped

The merge happens in GitHub. Sitearc only records a change as shipped when what merged is exactly what you approved, and it deployed.

Setup plans

A new Astro site or Google measurement starts as a plan. You approve it at a link, and the approval expires after 30 minutes.

Content

Approving a draft's preview approves that draft. Publishing opens a pull request, never a live post.

Results your agent can't choose.

Sitearc measures what shipped and computes the result itself. Your agent can only ask for it.

  • Changed pages are compared with holdout pages over the window your agent predicted.
  • The result is won, backfired or inconclusive. Inconclusive is a real result, not a failure.
  • Your agent logs confounders, such as an algorithm update. Only a person can resolve one.
  • Each learning carries a tier, from 1 for measured to 4 for inferred. Sitearc sets the tier.

Where the review step stops.

Know the edges before you rely on it. These are true today.

  • Sitearc never merges. It can't stop someone merging an unapproved pull request in GitHub. It just won't count it as shipped.
  • The preview audit verdict comes from your agent, not from a Sitearc check.
  • Some writes skip the dashboard: project context, form destinations, rank trackers and saved keywords.
  • Rank checks spend credits only within a limit you approved.
  • The WordPress tools would write to a live site with no preview. They need the WordPress plugin, which is not released.
  • Sitearc's plugin and skills for Claude Code and Codex are not public yet.
  • Claude on the web is tested end to end. The Claude Code and Codex steps are documented, not tested on every client.

Connect your agent today.

Add every site you run, free. Then point your agent at one address.

Start freeMCP reference