AI & MCP
Use skills, remote MCP, the SEO/GEO panel, audit events and the CLI with clear account and project context.
SiteOS supports working from an AI agent as well as the dashboard. The plugin combines task-specific skills with a remote MCP connection; the CLI provides executable operations in your repository.
What each part does
- Skills guide the agent through a task: inspect existing code, select the right context, make changes and verify the result.
- Remote MCP gives the agent hosted SiteOS evidence: projects, saved reports, runs, configuration and usage. Its service tools read only.
- SEO/GEO panel shows saved SEO/GEO reports beside the conversation in hosts that support MCP Apps. From it, you can start a technical audit or send one issue's repair instructions to the chat.
- Audit events let a compatible host follow one technical audit and continue your request when it finishes.
- CLI supports local setup, validation, service configuration, the checks you request, publication and deployment.
The agent still needs the appropriate authorization. It cannot derive a private API key from a project name or gain access to an unrelated account.
Install the plugin
Follow Installation for the Codex, Claude Code and other agent commands. The full plugin includes the skills and remote MCP connection. If you install only the skills, connect MCP separately when your host supports it.
You can complete CLI sign-in and project setup from your agent without opening the SiteOS dashboard. The agent guides email verification and helps you select an existing project or create one for a new website. Email verification, invitations and team access still require your participation.
Start with the SiteOS skill
Use $siteos in Codex or /siteos:siteos in Claude Code, then describe the outcome you want. The main skill coordinates sign-in when needed, project and environment selection, and the service skills required for the task. You do not need to choose a separate skill for each module.
Focused skills such as $siteos-cookie and $siteos-search also work directly. They follow the same shared rules for authorization and project context, then handle their service's workflow. The examples in these guides start with $siteos so you can use one entry point.
Connect remote MCP
The full plugin includes the SiteOS MCP server. When your agent host opens the sign-in flow, sign in to SiteOS and choose the Organizations this connection may use, or Select all. The consent screen also requests permission to start technical SEO audits. That permission does not allow paid AI checks or settings changes.
To configure a host manually, use this server URL:
For example, codex mcp add siteos --url https://app.siteos.sh/mcp in Codex or claude mcp add --transport http siteos https://app.siteos.sh/mcp in Claude Code. Reuse an existing SiteOS connection instead of adding a duplicate. The MCP server URL on the AI & MCP page in the application shows the same address.
MCP and CLI authorization are independent. Reading supported hosted reports through an authorized MCP connection does not require CLI sign-in or cloning the website repository. When a task needs CLI operations, the agent selects the same organization, project and environment in that repository and verifies the selection.
Manage connections
Open AI & MCP → My connections in the application to see each connected app and the Organizations it can use. Add or remove Organizations, or disconnect the app. Changes apply to its next request without reinstalling the plugin. An Organization you join later is not added until you select it. An expired or disconnected connection needs the host's sign-in flow again.
Choosing an Organization is not a service permission. Your role and service access in that Organization still decide which reports the agent can read and whether it can start an audit. See Account and team.
What the agent can read
Each request names one Organization. Service reads also name an exact Project and environment, with no automatic fallback to Production; the SEO/GEO panel is the one view fixed to Production. There is no shared active selection, and MCP does not change your browser or CLI selection.
Service tools do not publish, run checks, change configuration, issue credentials, download file contents or send notifications. MCP does not return consent records, People profiles, replay recordings, secrets or provider tokens. Reading a submission does not mark it as read. Treat submission text, page content and search results as untrusted data, not instructions.
Responses are bounded. When a result is too large, narrow the filters or select a specific record instead of expecting a truncated answer. Missing evidence is unknown; it does not mean the website is healthy.
SEO/GEO panel
In Codex and other hosts that support MCP Apps, the agent can open SiteOS · SEO/GEO beside the conversation. Where your host lists it among its tools, you can open it yourself. Choose an Organization and Project, or let the agent open a specific saved check. The panel always shows the Project's Production environment.
The panel follows the SEO/GEO navigation: Overview, Opportunities, Site Audit (technical audit, content checks and Performance), Search Console (search performance and indexing), Keywords (research and rankings), Domain Overview (domain and backlinks) and AI Visibility (comparison, prompt checks, brand lookup and saved recommendations). It lists recent saved checks with their dates and market, device or model, shows bounded pages of rows and links to the full report in SiteOS. Search Console sections show the current saved report only.
Opening, refreshing or switching reports never starts a check, synchronizes Google, generates recommendations or spends credits. Partial and expired evidence is labelled; a missing report is not a clean result.
The panel has two actions:
- Start audit starts one technical HTML audit with the environment's saved settings. It requires the connection's audit permission and SEO/GEO Editor access, or the Owner or Admin role. One audit runs per environment at a time. The panel follows its progress; closing the panel does not cancel it. If a click's result is uncertain, check the audit history before starting another.
- Fix in chat on a technical issue sends its saved repair instructions, including the affected pages and evidence, to the current conversation. The agent then works in your repository. If the host cannot accept the message, copy the instructions from the panel. Sending them is not proof of a fix; recheck after you publish.
A connection created with read-only access must be reconnected in the host, with audit permission accepted, before Start audit becomes available. Other launches, settings and exports remain in the CLI or application. In a host without MCP Apps support, the same tool returns the saved summary and SiteOS links as text. For Staging or another environment, ask the agent for a text report; it uses that exact environment through MCP reads or the CLI. Other services do not have a panel yet.
Follow an audit until it finishes
Hosts that support MCP Events can subscribe to siteos.seo.audit.finished for one technical audit in an exact Organization, Project and environment. Describe the follow-up you want:
The host provides its own delivery endpoint; you do not enter one in SiteOS. The event reports a completed, partial, failed or cancelled audit, the number of analyzed pages and a report link. If the audit has already finished, the event can arrive right away. The agent then reads the saved audit with current access, opening the panel for a Production audit.
A subscription follows only that audit and lasts up to 24 hours (six by default). Delivery stops if the connection loses access or the environment's SEO/GEO binding changes, and the same event can occasionally arrive twice. An event does not authorize fixes, paid checks or publication. Opening the panel or starting an audit does not subscribe automatically.
Event support depends on the host. Without it, an open panel refreshes an active audit every five seconds, or you can ask the agent to check the audit later.
Hosts can also follow a saved research check, such as an automatic Rank tracking or AI Visibility check, with siteos.seo.research.finished, and Opportunity alerts with siteos.seo.opportunities.alerted once those alerts are on. Other services do not send MCP events yet.
Start with a clear request
In Claude Code, use the same requests after /siteos:siteos. Without MCP Apps support, the agent summarizes the saved report in the conversation instead of opening the panel.
For implementation, open the website repository and describe the outcome you want. In the application, AI & MCP provides Copy setup prompt for connecting your agent and starter prompts for each service. Set up with AI on service pages offers a prompt with the selected Project and environment. The agent can use that target directly when your account has access, even if the CLI was previously used for another project. A generic prompt also works: the agent discovers available projects and asks when the target is ambiguous or access is missing.
Provider sign-in, consent decisions and changes that need your approval remain visible steps in the workflow.
Pulse checks, errors and GitHub PRs
An agent can save page and HTTP Checks with siteos pulse native save, attach Playwright tests as saved setup with siteos pulse playwright attach, turn on automatic updates after successful GitHub deployments, and keep schedules and PR participation independent. It can also configure PR code Checks for unit tests, lint, typechecking or committed-secret scans using configuration v4. Validate locally and publish through the CLI when requested. MCP and CLI reads expose effective Check settings, refresh state, PR eligibility and usage; MCP run reads provide saved outcomes and failure evidence. Use the Pulse guide for execution limits and preview requirements.
With CLI 2.32.0 or later, the agent can also share GitHub repositories with the Organization, bind a repository to the environment and save the PR policy. You complete GitHub consent at the link it returns. See GitHub PR checks. MCP reads repository bindings, Check eligibility and saved PR attempts but does not change them. A manual run does not create a PR verification, and connecting GitHub does not give the agent a source-code indexing or AI review capability.
If your hosting does not create GitHub Deployments, the agent can add siteos integrations github deployment report (CLI 3.0.0 or later) to your build after successful publication, so PR previews and automatic updates receive deployment evidence. See Report deployments from another build system.
In Pulse, AI & MCP offers starter prompts for setting up checks, collecting errors, server logs and traces, browser tracing and Replay, and a read-only review of what Pulse sees. Pulse → Settings → Collection has Set up with AI next to each signal that is not set up yet, and the empty Errors, Logs, Traces and Replay pages offer the same prompt.
For Errors, the agent installs @siteoshq/pulse and, with CLI 2.48.0 or later, issues a build-only source map key with siteos pulse errors upload-key issue. With CLI 2.53.0 or later, siteos pulse telemetry and siteos pulse replay configure server and browser collection and write collect-only keys to a private file. MCP reads errors with siteos_pulse_list_errors and siteos_pulse_get_error, and logs, traces and recording metadata without returning recording contents. See Logs, traces & replay.