BYO AI · Model Context Protocol · 8 supported clients
Bring your own AI client. Admaxxer ships a Model Context Protocol (MCP) server that any MCP-compatible AI can plug into — Claude Desktop, Claude Code, ChatGPT Desktop, Cursor, Windsurf, OpenClaw, Cline, Zed. Generate one token in Admaxxer, paste it into your client's config, and your AI can read every connected ad account, query revenue and attribution from your Tinybird pipes, and (with explicit confirmation) pause campaigns or change budgets.
Each guide below covers the exact config-file path for that client, the JSON snippet to paste, how to verify the connection, and how to troubleshoot when tools don't show up.
The Model Context Protocol is open — if your client speaks MCP, Admaxxer works. We ship one-click configs for the 8 most popular clients below.
Anthropic's desktop app — native MCP.
Use Claude Desktop on macOS or Windows. Paste the snippet, restart the app, and Claude can read your campaigns, revenue, and attribution.
Config location: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) · %APPDATA%\Claude\claude_desktop_config.json (Windows)
Claude in your terminal — works the same way.
Claude Code (Anthropic's CLI) reads MCP servers from `.claude/settings.local.json` or your user-level config. Paste, then `/mcp` to verify.
Config location: .claude/settings.local.json (project) · ~/.claude/settings.json (user)
OpenAI's desktop app — MCP via Settings.
ChatGPT Desktop ships MCP server support under Settings → MCP Servers. Paste the URL + Authorization header below.
Config location: ChatGPT Desktop → Settings → MCP Servers → Add
AI-first code editor — native MCP.
Cursor reads MCP servers from `.cursor/mcp.json` (project) or `~/.cursor/mcp.json` (global). Paste and reload Cursor.
Config location: .cursor/mcp.json (project) · ~/.cursor/mcp.json (global)
Codeium's agentic IDE — speaks MCP.
Windsurf reads MCP servers from `~/.codeium/windsurf/mcp_config.json`. Paste and restart Windsurf.
Config location: ~/.codeium/windsurf/mcp_config.json
The lobster way — open-source autonomous agent.
OpenClaw reads MCP servers from `~/.claw/config.json` under the `mcpServers` block — same shape as Claude Desktop. Paste and restart `claw`.
Config location: ~/.claw/config.json
VS Code AI agent — MCP-native.
Cline (the autonomous VS Code agent) reads MCP servers from its settings panel. Open Cline → MCP Servers → Edit MCP Settings, paste the snippet.
Config location: VS Code → Cline panel → MCP Servers → Edit MCP Settings
High-performance editor — MCP via context servers.
Zed surfaces MCP servers as 'context servers' under Assistant → Settings. Add an entry with the URL + bearer token below.
Config location: ~/.config/zed/settings.json (context_servers block)
Admaxxer's MCP server exposes 6 tools. Read tools run automatically; write tools (pause / update_campaign) require explicit user confirmation in Admaxxer.
admaxxer_list_connectionsreadLists every Meta + Google ad-platform connection in the workspace, with status (healthy / token expired / disabled), platform, and account label. Used by the AI to discover which connections to query for insights.
Example prompt: “What ad accounts do I have connected?”
admaxxer_list_campaignsreadLists active or paused campaigns for a given connection (Meta or Google). Returns campaign ID, name, objective, status, daily budget, and lifetime spend.
Example prompt: “Show me all my Meta campaigns sorted by spend.”
admaxxer_get_campaign_insightsreadPulls metrics for one campaign over a date range — impressions, clicks, spend, conversions, ROAS, CPA. Optional breakdowns by age, gender, device, region.
Example prompt: “How did campaign 'BFCM Scale' perform last week?”
admaxxer_get_account_insightsreadAccount-level rollup of insights across all campaigns for a connection. Optional breakdowns. Best for high-level ROAS / spend trends.
Example prompt: “What's my Meta account ROAS for the last 30 days?”
admaxxer_query_metricsreadRuns a whitelisted Tinybird analytics pipe (visitors, revenue, MER, LTV, MMM, attribution, forecast). The workspace_id is injected server-side — the AI cannot see another workspace's data.
Example prompt: “What's my blended MER for the last 14 days vs the prior 14 days?”
admaxxer_get_workspace_contextreadReturns the workspace's plan tier, currency, time zone, connected platforms, and last-sync timestamps. Useful for the AI to ground its answers.
Example prompt: “What plan am I on and when was my data last synced?”
No. MCP is included on every plan, including the free Pixel tier. Paid plans get higher rate limits on the Tinybird analytics pipes and the destructive-tool gate (pausing / updating campaigns).
No. The MCP server only sees the tool calls your AI makes (e.g. 'list_campaigns for connection X'). It does not receive your chat history, system prompts, or anything else from the client. We log the tool calls to your `ad_sync_logs` table for your own audit, but never the surrounding conversation.
Open Admaxxer → Settings → AI Providers → MCP and click 'Revoke' on the leaked token. We store only a SHA-256 hash, not the plaintext, so revocation is immediate and irreversible. Issue a new token and update your client config. The old token cannot be reused.
No. The MCP surface only exposes 4 read tools and 2 write tools (pause campaign, update campaign budget). Account-level destructive operations (delete account, transfer billing) are not in the tool catalog at all. Even the write tools require explicit confirmation in Admaxxer's UI before execution.
Any client that speaks Model Context Protocol. We ship one-click configs for Claude Desktop, Claude Code, ChatGPT Desktop, Cursor, Windsurf, OpenClaw, Cline, and Zed. The protocol is open — if your client supports MCP, Admaxxer works.
Tokens are scoped to a single workspace today (the one active when you issued the token). Per-tool scoping is on the roadmap — for now you can issue separate tokens per client and revoke any individually.
BYOK (bring-your-own-key) lets you use Admaxxer's built-in `/chat` with your own AI provider key (OpenAI, Anthropic, OpenRouter, etc.). MCP is the reverse — you use your own AI client (Claude Desktop, ChatGPT, Cursor) and connect it to Admaxxer's data. Both can be used together: connect Cursor via MCP, and have Admaxxer's `/chat` use your Anthropic key.
Generate your MCP token, paste it into your AI client's config, and you're ready to query your ad data conversationally. Tokens are per-workspace, revocable, and stored as SHA-256 hashes — we never see the plaintext after issuance.
Generate your MCP token