generated: '2026-09-13'
method: probed
source: >-
  https://docs.getmembrane.com/docs/ways-to-use-membrane/mcp.md (endpoint published by the provider),
  then probed live 2026-09-13
status: gated
server:
  name: membrane-integrate-anything
  transport: http
  url: https://api.getmembrane.com/mcp/integrate-anything
deployment:
  mode: both
  endpoint: https://api.getmembrane.com/mcp/integrate-anything
  install: git clone https://github.com/membranehq/mcp-server.git && npm install && npm run build
  package: https://github.com/membranehq/mcp-server
  auth: oauth
  verified: probed
  checked: '2026-09-13'
  note: >-
    mode is `both` because the provider ships two genuinely different things. The hosted endpoint at
    api.getmembrane.com/mcp/integrate-anything is callable by an agent right now - it answers an
    anonymous POST with a 401 and advertises its own OAuth metadata, which is a reachable agent
    surface. Separately, github.com/membranehq/mcp-server is an open-source server a customer runs
    themselves. It declares the npm name "integration-app-mcp-server" v1.3.0 in package.json but is
    NOT published to npm (registry 404 on 2026-09-13), so there is no `npx -y` line to record; the
    only published install path is a git clone or the repo's Dockerfile. Recording an npx command
    here would be an invention.
probes:
- url: https://api.getmembrane.com/mcp/integrate-anything
  method: POST
  body: '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'
  accept: application/json, text/event-stream
  http_status: 401
  body_snippet: >-
    {"message":"This request requires authentication, but access token was not provided",
    "type":"bad_request","key":"not_authenticated"}
  verdict: gated
- url: https://api.getmembrane.com/mcp
  method: POST
  http_status: 404
  verdict: not a server path; only the named /mcp/<skill> paths exist on the hosted service
- url: https://api.getmembrane.com/.well-known/oauth-protected-resource/mcp/integrate-anything
  method: GET
  http_status: 200
  verdict: RFC 9728 protected resource metadata served anonymously
authorization:
  model: oauth2
  spec: RFC 9728 protected resource metadata + RFC 8414 authorization server metadata
  resource: https://api.getmembrane.com/mcp/integrate-anything
  authorization_servers:
  - https://api.getmembrane.com
  token_type: tenant
  bearer_methods_supported:
  - header
  dynamic_client_registration: https://api.getmembrane.com/oauth/register
  pkce: S256
  alternatives_documented:
  - Bearer token in the Authorization header
  - token appended to the URL as a query parameter
  documents:
  - well-known/integration-app-oauth-protected-resource-mcp-integrate-anything.json
  - well-known/integration-app-oauth-authorization-server.json
tools_note: >-
  Anonymous tools/list is auth-gated (401), so no live inputSchema set was captured. The tool NAMES
  and their backing REST calls are nevertheless first-party and verifiable: the provider publishes
  tools/integrate-anything.ts in github.com/membranehq/agent-skills, which defines the same
  integrate-anything tool surface this endpoint is named for and shows the exact REST path each tool
  calls. Those 11 tools are recorded in mcp/integration-app-tool-crosswalk.yml, bound to verified
  operationIds in the published OpenAPI. They are NOT recorded here as a derived tool list.
  Membrane's hosted server is additionally dynamic: in static mode it returns every action available
  across a tenant's connected integrations, so the live tool set is tenant-specific and cannot be
  enumerated anonymously even with credentials to a different workspace.
modes:
  static: returns all available tools (actions) for every connected integration - the default
  dynamic: '?mode=dynamic returns a single `enable-tools` tool the agent calls to enable what it needs'
  source: https://github.com/membranehq/mcp-server#readme
transports:
  streamable_http: /mcp
  sse: /sse
  sse_status: deprecated by the provider, per its own README, following the MCP spec deprecation