MCP tools

Sitearc runs an MCP server at https://app.sitearc.ai/mcp. It gives your agent access to site context, research, feedback and the tools listed below. A person still approves proposals.

Connecting

For Claude Code, the plugin is the easy path. See the plugin doc.

For any MCP client, add the hosted server:

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

The OAuth flow

The server is an OAuth resource, and the flow is standard.

1. Your client hits /mcp with no token and gets invalid_token back.
2. It reads the resource metadata and finds the authorization server.
3. It registers itself dynamically, if it has not already.
4. You sign in at app.sitearc.ai and approve.
5. The client receives a token and refreshes it from then on.

Two scopes are requested. mcp grants tool access and is required. The offline_access scope is what lets the client refresh without sending you back to a browser.

Your grant is per client. Revoke it from your account and that client stops working, with no effect on the others.

Wire names

Inside Claude Code and most MCP clients, a tool is namespaced by its server. The wire names look like this:

mcp__sitearc__get_project_brief
mcp__sitearc__research_keywords
mcp__sitearc__ship_experiment

The short name is the tool’s own. The prefix is what your client adds.

The loop

Tool	Purpose
get_project_brief	The session’s starting point: connections, learnings, running experiments, confounders
get_budget_status	Remaining spend before anything metered is called
link_repo	Connect the project to the repository a change lands in
create_proposal	A change with evidence and a prediction, for a person to approve
list_proposals	The queue for a project
get_proposal	One proposal’s change, evidence and status
resolve_preview_url	The preview deployment a proposal’s pull request produced
ship_experiment	Ship an approved proposal and open its measurement window
list_experiments	What is running and when each window closes
get_experiment	One experiment, its prediction and its state
get_measurements	The collected rows for an experiment, cohort and holdout
conclude_experiment	Apply the standard and return a verdict
record_confounder	Anything else that moved during a window
list_confounders	Confounders recorded against a project
record_learning	Write a learning at the tier its evidence earns
search_learnings	What this project has already proved, with tiers

Projects and memory

Tool	Purpose
whoami	Who the token belongs to, and what it can reach
list_projects	Every site in the organisation
create_project	Add a site
get_project_context	Shared business memory: goal, positioning, key pages
update_project_context	Write that memory back

Keyword and SERP research

Tool	Purpose
research_keywords	Discover keyword ideas from seed topics
get_keyword_metrics	Volume, difficulty and intent for named keywords
save_keywords	Save and tag terms worth targeting
list_saved_keywords	The saved set for a project
get_serp_results	The live result page for a query
find_serp_competitors	Who else ranks across a keyword set
get_domain_overview	A domain’s organic footprint
get_domain_keyword_suggestions	Keywords a domain could plausibly target

Rank tracking

Tool	Purpose
create_rank_tracker	Set up tracking for a project
get_rank_tracker	Read a tracker and its latest positions
add_rank_tracking_keywords	Add terms to a tracker
remove_rank_tracking_keywords	Drop terms from a tracker
estimate_rank_tracker_cost	What a run will spend, before it spends it
run_rank_tracker	Run the tracker now
get_ranked_keywords	Everything a domain currently ranks for

Backlinks

Tool	Purpose
get_backlinks_overview	Referring domains, authority and totals
get_backlinks_profile	The individual links, with anchors and targets

Site audits

Tool	Purpose
run_site_audit	Start a technical crawl and audit
get_audit_status	Whether the audit has finished
get_audit_issues	The findings, grouped and prioritised
get_audit_pages	Per-page results from the audit

Content engine

Tool	Purpose
create_content_brief	Add a keyword opportunity to the backlog; a person approves it before anything is written
list_content_backlog	The briefs waiting on a person, with target keyword, volume and difficulty
generate_content_draft	One model call writes a draft for an approved brief; spends credits
get_content_document	One document in full: body, metadata, state, review notes, versions, change rounds
revise_content_document	A new version with a mechanical fix; stays in review
submit_content_for_approval	Put a verified draft in front of a person; approves nothing
request_content_changes	Send a document awaiting approval back with a numbered round of notes
publish_content	Publish an approved draft as a WordPress draft or a pull request; never live

Brand and compliance

Tool	Purpose
get_brand_profile	The versioned brand document: identity, vocabulary, colors and type by role, rules
propose_brand_patch	File a change to the brand document for a person to apply; changes nothing itself
get_compliance_profile	Banned words, banned claims, hedging rules and name styling; optionally check a text

Site crawl

Tool	Purpose
discover_urls	Every same-origin URL from robots.txt, the sitemaps and the home page navigation
fetch_page	One URL’s raw HTML and redirect chain; a JavaScript render spends credits
extract_page	What an SEO review reads from one page: title, meta, headings, body, links, images, schema
crawl_site	A breadth-first crawl bounded by the plan’s page cap; rendered pages spend credits

Recommendations

Tool	Purpose
list_recommendations	The project’s queue of drafted changes waiting on a person
get_recommendation	One recommendation in full: evidence with tiers, the change, the prediction, the cost
update_recommendation_status	Accept or dismiss; dismissing needs a reason, accepting creates a proposal
claim_work	Claim accepted or in-progress queue work for a short lease
renew_work_claim	Renew the lease before expiry; stop work if renewal fails
release_work_claim	Release a live claim without marking the work shipped or approved
submit_work_for_review	Submit structured work after a live claim; queues human review only
get_work_review	Read submitted revisions and human decisions for one recommendation
verify_work_revision	Save GitHub evidence for one human-approved revision
list_work_verifications	Read saved provider checks for one approved work revision

Submitted work keeps up to fifty revisions. Each artifact is capped at 128 KiB. submit_work_for_review needs a fresh requestId UUID for new work. Reuse the same requestId only when retrying the exact same artifact.

get_work_review may show currentDecision and prior decisions. Those are human decisions already recorded in the dashboard. The returned effects fields mean this tool did not approve, verify, deploy, or ship the work. They do not state whether some other path deployed it.

verify_work_revision records provider evidence only. It can read the PR head, merge state, merge commit, deployment state and preview match. It does not approve, merge, deploy, publish or ship anything.

list_work_verifications reads saved checks. It does not contact GitHub, use credits or change the work.

Feedback sessions

Tool	Purpose
create_review	Open a shareable feedback session over pages of a site and return its link
list_review_comments	Every comment on a session, public and internal, for triage
resolve_review_comment	Mark a comment resolved, optionally queueing it as a recommendation

Connected sites

Tool	Purpose
wp_list_posts	Posts and pages on the connected WordPress site, with status and modified date
wp_get_post	One post or page: title, body, excerpt, slug and SEO meta
wp_update_post	Rewrite the title, body, excerpt or slug of an existing post; never changes publication status
wp_create_draft	A new post or page saved as a draft for a person to review
wp_set_seo_meta	The SEO title and meta description, through the installed SEO plugin
wp_upload_media	Add an image to the media library, with alt text
github_open_pull_request	Open a pull request on the connected repository for a code change

Search Console and Analytics

Tool	Purpose
get_search_console_performance	Clicks, impressions, position and queries
inspect_urls	Google’s own index status for specific URLs
get_search_opportunities	Pages close to ranking, from your own data
get_google_analytics_organic_overview	Organic sessions and their shape
get_google_analytics_organic_landing_pages	Which pages organic traffic lands on
get_google_analytics_page_performance	Engagement per page
get_google_analytics_traffic_acquisition	Where the traffic came from
get_google_analytics_key_events	Conversions, as your property defines them
get_google_analytics_ecommerce_performance	Revenue and product performance
get_google_analytics_site_search	What people search for on your own site
get_google_analytics_audience_breakdown	Who is arriving, by dimension
get_google_analytics_measurement_health	Whether the property is measuring correctly

Local presence

Tool	Purpose
search_local_businesses	Find a business in Google’s local index
get_business_profile	The profile as Google holds it
get_business_reviews	Reviews on the profile
get_business_updates	Posts on the profile
get_google_business_questions	Questions asked on the profile
list_business_categories	The category taxonomy to choose from
get_local_serp_results	The local pack for a query and place
get_local_rank_grid	Maps visibility across a geography

Answer engine visibility

Tool	Purpose
get_llm_prompts	Questions assistants are answering in your topic
get_llm_visibility	Whether a site is cited in those answers

What the tools refuse

Refusals are in the code, not in a prompt. There are six on the shipping path.

1. A proposal that is not approved.
2. A proposal marked approved with no recorded decider.
3. A proposal that has already shipped.
4. A measurement window that disagrees with the prediction.
5. A failed preview audit.
6. No measurement source connected to read the result.

Two more sit earlier. create_proposal refuses a proposal carrying no evidence. It also refuses a new_page proposal outright.

And one is structural. There is no approval tool, and none can be added by accident. The module that records a decision is not importable from the agent-facing code. A test fails if it becomes so.

conclude_experiment has no field for a verdict. An agent cannot state its own outcome, only ask for one.

Rate limits

Requests per minute are counted per organisation, not per token.

Free allows thirty MCP requests per minute.

Pro has defined allowances.

The in-app chat is being built. Current free chat budget is eight tool calls per turn.

A capped turn should stop and say so.

Cost

Research tools spend credits. The server tells every agent to ask you before a planned batch over 2,000 credits.

get_budget_status is free and reports what is left.

Website forms

Tool	Purpose
list_forms	Find forms for a site before creating another
create_form	Save a field definition and return its public POST endpoint
get_form	Read the saved fields, allowed origin, status and retention
list_form_destinations	Read configured email and webhook destinations
create_form_destination	Add a future email or webhook notification destination
disable_form_destination	Stop future queued delivery to one destination
submit_test_form	Save a labeled test response and queue real test notifications
verify_form_receipt	Check that a real response was saved and read notification states

Install the returned endpoint in your website source, then submit a response through the website. Pass its receipt ID to verify_form_receipt. A saved configuration alone does not prove the form works.

submit_test_form checks backend validation and storage. It does not visit your website or check its installed form. Use synthetic contents and reuse the idempotency key when retrying. Configured destinations can receive real test notifications. Ask before sending a notification test. Webhook setup returns its signing secret once. Save it in the receiving service’s secret store, never in website source.

The verification result separates saved responses from notifications. provider_accepted means the destination accepted the request. It does not prove that an email reached a person’s inbox. Pending deliveries need another check after the background job runs.