MCP: Claude, ChatGPT, Cursor
The Skael MCP connects your AI assistant to Skael. Work in plain language: find companies, check websites, enrich contacts and build verified lead lists right in the chat. MCP is the Model Context Protocol, so every assistant that supports remote MCP servers can use it.
Status: beta, October 2026. The menus of Claude and ChatGPT change from time to time; if a label below differs from what you see, look for the closest match.
What it does
Your assistant gets a set of Skael tools. You describe what you need, the assistant picks the tools, and Skael does the work on your account:
- Lead lists. "Finde mir Hausverwaltungen in NRW, keine Makler" starts a short interview, a paid test on a small sample, and after your OK the full run. The CSV arrives by mail and on the run page.
- Single lookups. "Finde die Website zu Müller Haustechnik GmbH in Köln" finds and verifies the right domain against the imprint.
- Your account. "Was hat unser Team diesen Monat verbraucht?" reports usage, campaigns, prospects, members and API keys.
Every tool that costs credits tells you the price first. Nothing runs in the background without your confirmation.
Connect
The server address is the same for every client:
https://api.skael.de/mcp
You find the address, a copy button and these steps in the dashboard under Settings → MCP (/settings/mcp). Every way there takes three steps: copy the address or create a token, add it to your assistant, then Check the connection. The page waits for the first call and shows Connected when it arrives.
A step-by-step guide in German for ChatGPT, Claude, Grok, Le Chat, Cursor and VS Code, with example prompts, is at MCP Server verbinden.
Claude (claude.ai and Claude Desktop)
- Open Customize → Connectors, click + Add, then Add custom connector.
- Enter
Skael as the name and paste the server address.
- Under Authentication choose Sign in now (or Sign in when needed), sign in with your Skael account and pick the organization.
On Team and Enterprise plans an owner first adds the connector under Organization settings → Connectors; every member then connects on their own under Customize → Connectors. On the Free plan you can add one custom connector.
With a token instead of signing in. If the sign-in hangs or your Claude account does not allow one, click Create token in the Claude steps on /settings/mcp. In the Add custom connector dialog choose No sign in under Authentication and add a header under Request headers: name Authorization, value Bearer <YOUR_MCP_TOKEN>. Claude sends the value exactly as typed, so keep the word Bearer and the space. Request headers is a beta that Anthropic is rolling out gradually; without it, the Claude desktop app works through Settings → Developer → Edit Config with this block in claude_desktop_config.json (needs Node.js, restart the app fully afterwards):
{
"mcpServers": {
"skael": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://api.skael.de/mcp", "--header", "Authorization:${SKAEL_AUTH}"],
"env": {
"SKAEL_AUTH": "Bearer <YOUR_MCP_TOKEN>"
}
}
}
}
ChatGPT
- Open Plugins, click the plus button and choose Add custom MCP server.
- Under Connection paste the server address and choose OAuth as authentication.
- Sign in with your Skael account and pick the organization.
As of 6 November 2026, the OpenAI Help Center documents full MCP access with custom servers for ChatGPT Business and Enterprise/Edu on the web; Pro can connect custom MCP servers with read and fetch permissions only. In a workspace the admin controls it with the Create plugins with MCPs permission. Where Plugins is missing, the older path still works: admins via Workspace settings → Apps → Create, permitted users via their own settings under Apps → Create.
ChatGPT offers OAuth or no authentication for custom servers and has no field for a fixed token. Skael needs a sign-in, so choose OAuth; if it hangs, use Claude, Cursor or Claude Code with a token.
Cursor
The Open in Cursor button on /settings/mcp installs Skael with one click. To add it by hand, put this into ~/.cursor/mcp.json:
{
"mcpServers": {
"skael": {
"url": "https://api.skael.de/mcp"
}
}
}
On the first call Cursor opens the Skael sign-in.
With a token. A token goes through mcp-remote (needs Node.js). Click Create token in the Cursor steps on /settings/mcp, replace the skael entry in ~/.cursor/mcp.json with the block shown there (the same block as for the Claude desktop app above) and restart Cursor.
VS Code
Run Chat: Open Customizations from the Command Palette, choose MCP Servers, then Add Server, pick HTTP and paste the address. Sign in with your Skael account when VS Code asks.
VS Code marks .vscode/mcp.json as deprecated but still reads it. If you use it, the top-level key is servers:
{
"servers": {
"skael": {
"type": "http",
"url": "https://api.skael.de/mcp"
}
}
}
Grok
- Open
grok.com/connectors, click New Connector, then Custom.
- Paste the server address and complete the sign-in with your Skael account. Grok then reads the Skael tools.
On Business and Enterprise a team admin first adds the connector in the xAI cloud console (+ Add Connector, then Other, paste the address, complete the sign-in); members can use it after that.
- Open Connectors, add a connector and choose the Custom MCP Connector tab.
- Enter
Skael as the name, paste the server address, optionally a description, and click Connect.
- Vibe Work detects the sign-in method itself; sign in with your Skael account.
Only an administrator can add custom connectors; on Free, Pro and Student plans the account owner is the administrator.
Claude Code and the Code tab of Claude Desktop
Open Settings → MCP (/settings/mcp), choose Claude Code or Claude Desktop: Code tab and click Create token. Skael creates a personal MCP token for you and shows a ready command:
claude mcp add --scope user --transport http skael https://api.skael.de/mcp --header "Authorization: Bearer <YOUR_MCP_TOKEN>"
Paste it into your terminal (macOS: Terminal, Windows: PowerShell), not into the Claude chat, and restart Claude Code or the Claude app. No /mcp, no Authenticate: this also works in the Code tab of Claude Desktop, which cannot open the OAuth sign-in. The page shows Connected as soon as the first call arrives.
If claude is not found, install the Claude CLI (irm https://claude.ai/install.ps1 | iex on Windows, curl -fsSL https://claude.ai/install.sh | bash on macOS and Linux) and make sure ~/.local/bin (Windows: %USERPROFILE%\.local\bin) is on your path. If the command answers that skael already exists, run claude mcp remove --scope user skael and try again.
Authentication
OAuth is the default. Your assistant opens the Skael sign-in, you log in with your normal account and choose the organization. Access tokens last 8 hours and refresh on their own for 60 days. You can see and end connections in your assistant's settings.
Personal MCP token. Every member creates one on /settings/mcp, right in the steps for Claude, Claude Code, Cursor or Token in a header (for your own agents, servers and automations). It goes into the Authorization header as Bearer eak_…. Skael also accepts the token without Bearer , in quotes, or as an X-API-Key header. It works only with the MCP (the REST API refuses it), belongs to one organization and acts with your own role there, exactly like an OAuth sign-in. You find your tokens on the same page and revoke them there; admins also see them in the API key list.
Organization API key. If access should not depend on one person, an organization admin creates an API key with the option Use for MCP on the API Keys page and sends it the same way. A key acts as the person who created it, inside the key's organization only, and that person must still be a member there.
Never put a token or key into the URL or the query string.
When it does not connect. Skael answers a refused call with a message that says why: no credentials at all, the placeholder from the instructions still in the header, a token Skael does not know or that was revoked, or a key without MCP access. Clients that show the server's answer (Claude Code with --debug, mcp-remote in its log) show that message.
Organizations
If you belong to one organization with the beta enabled, Skael works there. If you belong to several, Skael uses the organization you picked at sign-in. Otherwise the assistant asks you which one and passes its slug as the organization argument. With an API key, the key's organization is always used.
Members and admins can both connect. Admin-only tools (API keys, members, budgets) only appear for admins.
Costs and limits
- Tools use your normal credits. Every quote shows credits and the euro amount.
- Anything above 500 credits needs your explicit confirmation in the chat first.
- Daily caps over a rolling 24 hours: without a cap of its own the daily budget follows your balance, at least 1,000,000 credits. Admins can set a different budget per member.
- A lead list run has its own budget cap that you confirm before it starts; the run stops when it reaches it.
- Some tools (AI generation, website search) cost us cash per call. They need purchased credits or a paid plan.
Your assistant sees 15 tools at first:
whoami: who you are, your organizations and your balance.
start_lead_project, update_icp, estimate_market, draft_pipeline: describe the target group, answer the follow-up questions and see how big the market is.
plan_pipeline_test, confirm_pipeline_test: a paid test on a small sample, with examples and costs.
review_sample: go through the test results and say what was wrong.
plan_pipeline_run, confirm_pipeline_run, get_run_status, get_results: the full run and its delivery. get_run_status waits up to 45 s per call; the assistant keeps calling it itself until the run is done.
export_run (hidden, called through run_endpoint with tool: export_run): the finished result file itself (CSV text or Excel as base64, all columns and sheets, in parts for large files) plus a download link valid for 30 days. Works for tests and full runs. With mode: raw it returns the raw data of a run in any state, also while it runs or after it failed or was cancelled: one line per company with its end state and one line per company and step with status, attempts, credits, last error and the step's raw result as JSON. Free, for debugging.
find_company_domain: the verified website for a company name.
search_tools, run_endpoint: everything else.
More tools are found with search_tools and run with run_endpoint: usage reports, SkaelReach campaigns and prospects, API keys, members, budgets, documentation, pricing and every enrichment endpoint.
SkaelReach video campaigns run through hidden tools as well: list_campaigns, get_campaign, create_campaign, update_campaign, pause_campaign, resume_campaign, get_campaign_upload_url and complete_campaign_upload for the base video (get the address, PUT the file, complete the upload), get_campaign_viewers, list_prospects and add_prospects. Creating, editing, uploading, pausing and resuming are free; credits are charged per video that add_prospects creates.
LinkedIn data runs through four hidden tools (found with search_tools, run with run_endpoint):
search_linkedin_people: people by job title, company, location, industry and company size. 1 credit per person returned, limit up to 50; the next page comes with cursor.
find_company_employees: the employees of a company by domain or LinkedIn company page, filtered by level, function and country. 1 credit per person returned; the next page comes with page.
get_linkedin_person: the full profile of one LinkedIn URL (career, education, skills). 10 credits per profile found.
get_linkedin_company: industry, size, founding year and headquarters of a company by domain or LinkedIn page. 10 credits per company found.
None of them return email addresses or phone numbers, and a lookup without a result costs nothing. For contact data use find_email_linkedin or find_phone_linkedin with the person's linkedin_url.
TikTok and YouTube sourcing runs through eight hidden tools (found with search_tools, run with run_endpoint). Start with the lead finder of the platform; the other tools look at single accounts or pages:
find_tiktok_leads: a TikTok lead list from accounts of your niche (seed_usernames) and/or keywords, inside a follower band, optionally only with contact data. It snowballs over the accounts the seeds and the found leads follow: on TikTok that brings 23 to 27 leads per 100 lookups, the user search only 1 to 3. 2 credits per lookup, capped by max_credits (default 100). max_leads caps the number of leads (default 25, up to 200; limit works the same).
search_tiktok_users, get_tiktok_profile (bio link, Instagram handle, email from the bio; 3 credits per profile found), get_tiktok_following (the snowball step by hand). Pages cost 2 credits when they hold accounts.
find_youtube_leads: a YouTube lead list from keywords inside a subscriber band. Each channel in the band is read once with its links and the email the creator lists in the description; the business email behind YouTube's View email address button is not read. 2 credits per lookup, capped by max_credits. max_leads caps the number of leads (default 25, up to 200; limit works the same).
search_youtube_channels, get_youtube_channel (3 credits per channel found), get_youtube_channel_videos. Pages cost 2 credits when they hold results.
A lookup without a result costs nothing outside the lead finders. Results are cached for 30 days (lists for one day); the price stays the same.
Instagram data runs through fifteen hidden tools (found with search_tools, run with run_endpoint):
find_instagram_leads: the recommended entry point for leads from Instagram (businesses, creators, photographers, coaches, shops). Sources itself from seed accounts (similar accounts and followings), hashtag authors and keyword searches, then snowballs over similar accounts, and returns only accounts with a public email or phone. Always confirmed first (CONFIRMATION_REQUIRED with the maximum price), then runs async; fetch the result with get_enrichment. 50 credits per delivered lead, reserved max_leads × the lead price, capped by max_credits. max_leads caps the number of leads (default 50, up to 500).
get_instagram_profile: one profile by handle or user_id — name, bio, website, public email and phone when the account shows them, category, location, country, followers. 2 credits live, 1 credit from the cache (up to 30 days, steer with max_age_days), 0 without a hit.
search_instagram_accounts: accounts matching a keyword — the first step when there is no seed account yet. 2 credits per page, 1 cached, 0 for an empty page; the next page comes with cursor.
find_similar_instagram_accounts: the accounts Instagram suggests as similar to one account — the best step from a good hit to more of them. 2 credits, 1 cached.
list_instagram_hashtag_authors: the accounts posting under a hashtag (top or recent), deduplicated. 2 credits per page, 1 cached, paged with cursor.
list_instagram_location_authors: the accounts posting at a place (place name or location_id). 2 credits per page, 1 cached, paged with cursor.
list_instagram_hashtag_posts: the posts under a hashtag (top or recent) with likes, comments, views on reels and videos, author handle, time, type and a shortened caption. 2 credits per page, 1 cached, paged with cursor.
list_instagram_account_posts: the latest posts of one account with the same fields, plus the account's average likes and comments and its engagement rate over the page. Private accounts return nothing and cost nothing. 2 credits per page, 1 cached, paged with cursor.
get_instagram_post: one post by link, shortcode or media id with likes, comments, views or plays on reels and videos, type, time, author, full caption, tagged accounts and hashtags. Every call is a new measurement with measured_at; max_age_hours (0 to 24) steers how old a cached measurement may be. Reach and impressions are not public. 2 credits live, 1 cached, 0 when the post does not exist or cannot be read.
list_instagram_post_likers: the accounts that liked a post. Instagram shows only a sample of the likers; the total is like_count of get_instagram_post. 2 credits, 1 cached, 0 without likers.
list_instagram_post_comments: the comments of a post with the commenting account, the text cut at 300 characters, the time and the likes. 2 credits per page, 1 cached, paged with cursor.
list_instagram_tagged_posts: the posts that tag an account, whoever published them, with the same fields as the account posts, no stats. Private accounts return nothing and cost nothing. 2 credits per page, 1 cached, paged with cursor.
list_instagram_account_reels: the reels of one account, newest first, with plays plus the same stats as the account posts. 2 credits per page, 1 cached, paged with cursor.
list_instagram_following: the accounts an account follows — industry hubs and associations often follow many fitting businesses. Private accounts return nothing and cost nothing. 2 credits per page, 1 cached, paged with cursor.
list_instagram_followers: the followers of an account — mostly private persons, better for audience analysis than for leads. 2 credits per page, 1 cached, paged with cursor.
The list tools carry only short profiles without contact data; get_instagram_profile adds the public contact of one account, find_instagram_leads delivers finished leads with email or phone.
Support works from inside the chat. report_problem opens a ticket with Skael and answers with its number (SKP-…). wait_for_support_reply waits up to 30 s per call for our answer and returns as soon as it is there; on pending the assistant calls it again by itself. get_support_conversations lists the support conversations of your organization, including the ones from the chat window in the dashboard, and reads one in full. All three run through run_endpoint and cost nothing. Our answer also reaches you by email and in the dashboard chat.
Planning a large purchase from another provider? Your assistant first asks whether it should request a special price from Skael API, and only after your yes sends request_special_price with the volume, your use case and, if you name it, the other offer. It reaches our support team as a ticket (SKP-…), the answer comes by email and in the dashboard chat, and the assistant never promises a price itself. Comparisons with another tool are made only on a measured sample, and the result is reported as measured.
Admins manage members through hidden tools as well: list_members, plan_member_invite, plan_member_update (make someone an admin or a member) and plan_member_remove, each carried out by its confirm_* half after you said yes. Nobody changes their own role or removes the last admin, the same rule as in the dashboard.
list_invoices lists your organization's invoices with date, amount and status, optionally for one year, and gives a PDF link per paid invoice that works for 24 hours without signing in. Card details never appear in the chat. Members and admins can use it.
With the German registry pilot enabled, create_registry_monitor watches a company in the Handelsregister and reports every change to your webhook (50 credits, always confirmed first; the signing secret is shown once), and plan_registry_monitor_delete stops a watch. Transparenzregister credentials never go through the chat.
run_workflow starts one of your organization's workflows with the same checks as POST /api/workflows/{type}; called without workflow it lists the workflows you can start, with their required fields and cost. Anything above 500 credits is confirmed first.
Lead lists
- Interview. You describe who you are looking for. Skael asks what you do not want (brokers, chains, large companies) and which details matter.
- Market. Skael estimates how many companies match, by region.
- Test. A small paid sample shows real rows, what was dropped and why, and the cost per kept lead.
- Run. After your OK the full list runs within the budget you confirmed. For Google Maps you choose the search depth: until target starts with the largest cities and stops starting new companies once your target is reached; full searches every city above a minimum population (like a Clay table built from
find_cities × keyword) and stops only when every search is done or the budget is spent. The plan shows how many cities × keywords are searched and what both options cost; a search without hits costs nothing.
- Delivery. The CSV or Excel file comes by mail and stays on the run page under Leadlisten.
Playbook and prompts
You do not need a template. The MCP brings our own way of working: your assistant reads it when it connects, and get_guide returns the full playbook (Leitfaden: Leadlisten mit Claude) or a ready CLAUDE.md for one customer's folder. Clients that show MCP prompts offer them as shortcuts:
- Leadliste bauen: a lead list step by step, with costs up front.
- Kundenprojekt anlegen oder weitermachen: one lead project per end customer. Projects live at Skael, not in the chat, so "Mach bei Kunde X weiter" works in any new chat (
list_projects with suche).
- Clay-Workbook nachbauen: translates the columns of a Clay workbook into a lead project.
- Leitfaden: so arbeitet Skael: puts the whole playbook into the conversation.
Local mode: your own scripts
Power users who run their own scripts with Claude Code or Cursor can let the assistant prepare everything and run the work on their machine:
get_code_recipe returns a ready, documented Python script for one endpoint (batch or one request per row). It resumes after an interruption and never pays twice, because every request carries an Idempotency-Key.
create_upload returns a signed upload address for a CSV or XLSX file and the csv_url to use afterwards in a lead list, a match list or the batch endpoint.
Both are found with search_tools and run with run_endpoint. Details on the batch endpoint and the Idempotency-Key header: Lokal mit eigenen Skripten.
Beta
The MCP is in beta and enabled per organization. If your organization does not have it yet, or something does not work as described, write to support@skael.de.
German overview
Overview page in German: MCP Server für B2B Leads.