WundertreOS

MCP server overview

Connect AI agents to WundertreOS over the Model Context Protocol.

WundertreOS ships a first-class MCP (Model Context Protocol) server, so AI agents — Claude, ChatGPT, Claude Code, or anything else that speaks MCP — can operate your workspace directly: look up contacts, move deals, schedule social posts, pull ad performance, build workflows, and much more.

While the REST API covers contacts, deals, activities, and webhooks, the MCP server exposes the full WundertreOS tool registry — 200+ tools spanning CRM, prospecting, social, ads, email/SMS marketing, content generation, websites and landing pages, support, analytics, workflows, and business knowledge. These are the same tools that power Wunder, the in-app agent, so agent capabilities stay identical inside and outside the product.

Endpoints

Use caseURL
OAuth clients (claude.ai, ChatGPT, Claude Code)https://os.wundertre.com/mcp
API-key clients (scripts, header-capable MCP clients)https://api.wundertreos.com/functions/v1/api-mcp

The os.wundertre.com/mcp URL fronts the same server and additionally serves the OAuth discovery metadata MCP hosts need (/.well-known/oauth-authorization-server, /.well-known/oauth-protected-resource), so hosts can bootstrap the entire auth flow from that one URL.

Transport & protocol

  • Streamable HTTP in stateless JSON mode — every call is a self-contained POST; no session to manage.
  • Protocol versions 2025-06-18, 2025-03-26, and 2024-11-05 are accepted.
  • Supported methods: initialize, ping, tools/list, tools/call.

Authentication & scopes

The MCP server uses the same credentials as the REST API:

  • Authorization: Bearer <OAuth access token> — issued via the standard OAuth flow (what claude.ai and ChatGPT use, via automatic dynamic client registration).
  • Authorization: Bearer <API key> or X-API-Key: <key> — from Settings → Integrations & API.

Every tool is mapped to a scope family. tools/list returns only the tools your credential's scopes allow, and the scope is re-checked on every tools/call — a narrowly scoped connection simply never sees tools it can't use.

Read-only vs. action tools

Tools are annotated so MCP hosts can show the right approval UX:

  • Read tools (readOnlyHint: true) are safe to call freely.
  • Action tools create, update, send, or delete real data. They take effect immediately — the approval step belongs to the MCP host (e.g. Claude asks you before running them).
  • Destructive tools (deletes, cancellations) additionally carry destructiveHint: true.

Every action call is recorded in your organization's audit log with the tool name and arguments, attributed to "External agent (MCP)" — so there is always a trail of what an agent changed.

Agents act with your credential's permissions

A full-grant API key lets a connected agent do everything the key allows, without per-call server-side confirmation. Scope credentials to what the agent actually needs, and revoke them from Settings → Integrations & API if anything looks wrong.

Rate limits & response size

MCP calls share the org's rate limit (60 requests / 60 seconds) with the REST API — a busy agent conversation and your Zapier polling draw from the same bucket. Tool results are returned as text (plus structured JSON where available) and truncated at 100,000 characters.

What's next

  • Connect an AI agent — claude.ai, Claude Code, ChatGPT, and API-key clients
  • Scopes — the permission families that gate tools
PreviousActionsNext Connect an agent