MCP Setup

​

Prerequisites

• Sight AI team owner or admin
• An MCP client that speaks Streamable HTTP and OAuth 2.1 with PKCE: Claude (web, desktop, mobile, Claude Code), ChatGPT, Cursor, VS Code, Windsurf, Gemini CLI and most others

​

MCP server

https://app.trysight.ai/mcp

​

Connect from Claude

​

Connect from ChatGPT

​

OAuth endpoints

​

Register a client

• Client ID Metadata Documents (the MCP spec’s preferred mechanism): Claude and ChatGPT identify themselves with an HTTPS URL as their client_id; the server fetches that document, checks it and uses its redirect URIs. The consent screen names the domain the document came from.
• Dynamic Client Registration (RFC 7591): Cursor, Claude Desktop through mcp-remote, VS Code and others register a public client automatically when you connect.

1. Sign in as a team owner or admin.
2. Go to Integrations → Sight AI API → OAuth apps.
3. Click Create OAuth client, give it a recognizable name, and add a loopback redirect URI such as http://127.0.0.1/oauth/callback (or cursor://anysphere.cursor-mcp/oauth/callback for Cursor). Leave Server-side app unchecked — MCP clients are public clients that authenticate with PKCE only.
4. Copy the generated client_id if your MCP client requires a static client (see below).

​

Authorize

{
  "mcpServers": {
    "sight-ai": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://app.trysight.ai/mcp",
        "--static-oauth-client-info",
        "{\"client_id\":\"YOUR_CLIENT_ID\"}"
      ]
    }
  }
}

​

Available MCP tools (v1)

​

How the server behaves

• Instructions. initialize returns an instructions string your client gives the model: start with list_sites, which calls cost budget or credits, and that writes need the user’s go-ahead.
• Titles. Every tool carries a title (for example get_seo_rankings is “Tracked rankings”), which clients show in their tool lists and permission screens.
• Annotations. Every tool in tools/list carries readOnlyHint, destructiveHint, idempotentHint and openWorldHint, so clients can run reads without asking and confirm writes. Paid lookups (research_keywords, fetch_keyword_overview, explore_domain, suggest_seo_competitors, and the AI recommendations recommend_article_types / recommend_article_category) are not marked read-only: an uncached call spends the site’s daily lookup budget or the team’s credits. Writes that overwrite, delete, cancel or send email are marked destructive, as is article generation (it spends credits); additive writes that cost nothing (create a draft, add keywords, submit for indexing) are not.
• Permissions. A tool the connection’s scopes do not cover answers 403 with WWW-Authenticate: Bearer error="insufficient_scope" and the full scope list; clients that follow the MCP step-up flow (Claude does) re-run consent and retry the call.
• Validation. Arguments are checked against the tool’s schema before it runs: a missing siteId, an action outside the enum or a value of the wrong kind returns isError with the argument named and code: VALIDATION_ERROR.
• Errors. A tool that fails returns a normal result with isError: true and text such as The site's daily lookup budget is used up (code: BUDGET_EXHAUSTED), so the model can read the reason and tell the user. The codes match the REST error codes. Unexpected server failures return Internal error. Calling a tool that does not exist is a JSON-RPC -32602 error.
• Image generation is an app and agent feature; the MCP server does not offer generate_article_image. set_article_main_image sets a cover image from a URL.
• Single-site access. When the credential is limited to one site, siteId can be left out and that site is used.
• Plans. tools/list only lists the tools the team’s plan includes.

​

Sites & analytics (analytics:read)

• list_sites / get_site
• get_search_overview / get_search_timeseries / get_search_queries / get_search_pages / get_search_article_performance
• get_ai_visibility_summary / get_ai_visibility_trends / get_ai_visibility_mentions / get_ai_visibility_citations / get_ai_visibility_competitors

​

SEO module (analytics:read / seo:write, requires the SEO module)

​

Opportunities (opportunities:read)

• list_opportunities — filter by source (ai/search/all) and status (e.g. pending)
• get_opportunities_readiness — pipeline warmup status

​

Articles

• list_articles — filter by status (draft, generating, ready, failed)
• get_article — full article body, SEO fields, target keyword, and status
• get_article_limits — the team’s AI credit balance and usage this billing period (there are no per-article quotas)
• get_article_images — list hero/thumbnail/inline images for an article
• fetch_url_content — crawl a public URL (title, headings, body) for rewrite workflows
• get_writing_guide — pipeline writing playbook for listicle / step-by-step / explainer / comparison + business context
• recommend_article_types — recommend an article type per keyword
• recommend_article_category — recommend a CMS category from keyword/title
• list_article_authors — list Webflow/WordPress authors for batch assignment
• project_interlinks — estimate interlink-automation output for a given cron + end date

• generate_article — start async generation from a keyword (poll get_article until ready). Articles generated over MCP get no AI cover image: open the article in Sight AI (Content → Articles) and use Generate image there, or set one from a URL with set_article_main_image
• create_draft_article — create a blank draft to fill with update_article or edit_article_section
• update_article — edit title, HTML body, SEO fields, slug, and summary (status is not exposed)
• edit_article_section — surgical body edits: append/replace a section or rewrite the intro
• refresh_seo_title / refresh_seo_meta — AI SEO rewrites (rate-limited)
• set_article_main_image — set/clear the hero image from a URL
• create_article_batch — queue 1–100 articles through the full pipeline (author round-robin, categories, interlinking focus); each article spends AI credits. Like generate_article, no AI cover images over MCP

# Rewrite an existing page
fetch_url_content → get_writing_guide → create_draft_article → edit_article_section → update_article → sync_article

# Plan a batch
recommend_article_types → list_article_authors → create_article_batch

​

Keywords (keywords:read / keywords:write)

• list_keywords — paginated list with pipeline status; filter status: no_article for generation candidates
• get_keyword — lookup by keywordId, exact keyword text, or slug
• get_keywords_summary — totals, counts by status, and planning recommendations
• check_keywords — exact coverage check for up to 50 candidates against the full keyword list and existing articles; call it before recommending a keyword
• list_keyword_groups — the site’s keyword groups with ids and counts

• add_keywords — single keyword or bulk (comma/newline, max 100); dedupes existing slugs
• manage_keywords — add (same as add_keywords, with groupName/groupId), assign_group (group keywords that already exist; the group is created if missing) and remove (keyword ids, keyword texts, or every keyword a filter covers: status such as no_article, groupId/groupName, source, search; dryRun: true counts and samples first). Reads and writes are separate tools on the MCP server.

get_keywords_summary → list_keywords (status: no_article) → add_keywords
  → generate_article → get_article (poll) → update_article → sync_article

​

Agents & Automations (agents:read / agents:write)

• list_agents / list_agent_runs
• update_agent — toggle or tune an agent
• run_agent — on-demand run (rate-limited, agent must be active). Over MCP this is one tool per agent: run_search_opportunity_agent, run_ai_opportunity_agent, run_article_boost_agent, run_interlinking_agent, run_site_performance_agent, run_outreach_agent

​

CMS & integrations (integrations:read / integrations:write)

• list_integrations — connected CMS/platform integrations and which one is active
• get_cms_status — active CMS, sync eligibility, and manual-only integrations (Framer)
• get_article_sync_status — per-platform sync state for an article
• sync_article — push a ready article to the active CMS (WordPress, Webflow, Wix, Shopify, HubSpot, webhook)

​

Planner (planner:read / planner:write)

• get_planner — read the planner: view is settings, scheduled (the CMS publish queue, each item with origin autopilot|manual), generation (upcoming Autopilot slots) or autopilot (health); startDate/endDate narrow the queue views
• schedule_article — queue ready articles for a future CMS publish date
• manage_planner — change the planner
  • Actions: update_settings, update_scheduled, remove_scheduled, clear_scheduled, clear_generation, plan_generation

​

Indexing & setup health (indexing:read / indexing:write)

• submit_sitemap_gsc — submit the configured sitemap URL to Google Search Console
• get_gsc_status — GSC connection, property, and sync state
• submit_article_indexnow — submit one article URL to IndexNow (after CMS sync)
• get_indexnow_status — IndexNow verification + optional URL lookup
• get_account_health — composite setup troubleshooting (CMS, GSC, IndexNow, autopilot, blockers)

​

Outreach inbox (inbox:read / inbox:write)

• list_inbox — conversation list with latest preview; filter by status (unread, open, awaiting_human, etc.)
• read_message — full conversation thread + opportunity context + agent reasoning
• get_inbox_unread_count — unread thread/total counts
• mark_conversation_read — clear unread state on a thread
• reply_to_message — send a human reply (requires Outreach email setup)
• update_conversation_status — resolve, decline, reopen, pause, or resume AI
• take_over_conversation — hand awaiting_human threads to the Outreach Agent

​

End-to-end content workflow

list_sites → get_keywords_summary → list_keywords (status: no_article)
  → add_keywords (optional) → get_article_limits → generate_article
  → get_article (poll) → update_article (optional)
  → refresh_seo_title (optional) → get_cms_status → sync_article

list_sites → get_opportunities_readiness → list_opportunities
  → get_article_limits → generate_article → get_article (poll)
  → refresh_seo_title (optional) → get_cms_status → sync_article

list_agents → run_search_opportunity_agent (MCP; `run_agent` on REST) → list_agent_runs → sync_article