# NexoFlow MCP server

Source: https://nexoflow.ai/mcp

Remote Model Context Protocol server at https://nexoflow.net/mcp- connect Claude, Cursor, or any MCP client to real marketing data, with 29 scoped tools and no instant-publish path.

## What it is

A remote MCP server hosted by NexoFlow at https://nexoflow.net/mcp. Any AI assistant that speaks the Model Context Protocol- Claude Desktop, Claude Code, Cursor, and others- can read a brand's marketing data and create work inside NexoFlow without a browser, a scraper, or a custom integration.

Transport is streamable HTTP with JSON responses- deliberately not long-lived SSE- and each request is stateless. There is no local process, Docker container, or desktop agent to run.

## Three steps to connect

1. Point an MCP client at the hosted endpoint using a personal token or OAuth.
2. Grant scopes and projects deliberately. Read-only tokens never see write tools; write tokens cannot one-click grant across every client project.
3. Ask in plain language. The assistant reads Search Console, calendars, and performance, then drafts and schedules into NexoFlow- nothing goes live instantly.

```json
{
    "mcpServers": {
      "nexoflow": {
        "url": "https://nexoflow.net/mcp",
        "headers": { "Authorization": "Bearer nxf_mcp_…" }
      }
    }
  }
```

## Authentication

- Personal tokens: Account Settings → MCP. Name the credential, tick permissions and projects, set expiry (1-365 days, default 90). The nxf_mcp_… token is shown once; only a SHA-256 digest is stored, so NexoFlow cannot read the secret back.
- OAuth 2.1: full authorization server with discovery, mandatory PKCE (S256), refresh-token rotation, and an allowlisted redirect set. The consent screen forces project selection and labels write scopes "read and change".
- Rotation takes one click and leaves a 60-minute grace window.
- An active NexoFlow subscription is checked on every request; banned or unverified accounts are rejected.

## Tool catalog

29 tools: 18 read, 11 write. Search Console reads the same six-hour cache as NexoFlow's own content engine, so assistant suggestions and in-app generation share identical demand data.

Read tools include:

- nexoflow_get_search_keywords- top queries and striking-distance opportunities
- nexoflow_get_content_calendar- website, social, and Google Business items with gap detection
- nexoflow_get_winning_patterns- caption, CTA, format, and hashtag patterns that worked
- nexoflow_get_best_posting_times- day and hour this brand earns engagement
- nexoflow_get_search_performance- clicks, impressions, CTR, position with deltas
- nexoflow_get_ai_usage- generation counts, tokens, estimated spend

Write tools include:

- nexoflow_draft_website_post- draft a website article
- nexoflow_propose_blog_plans- propose plans, always unscheduled
- nexoflow_propose_social_drafts- generate social drafts
- nexoflow_schedule_website_post- schedule with a cancel window
- nexoflow_revert_write_operation- restore a previous write
- nexoflow_pause_autopilot- pause automated generation

## Scopes

Nine scopes, nothing inherited by accident. New credentials default to project discovery only.

- projects:read- projects and connection health
- content:read- plans, calendar, SEO fields
- content:read_full- full article text (separate on purpose)
- analytics:read- analytics and Search Console
- reviews:read- Google Business reviews
- content:draft- create drafts and ideas
- content:write- edit posts and plans
- content:schedule- schedule posts
- connections:write- pause connections or autopilot

Connections can be paused but never re-enabled over MCP. Deletes, organization invites, and API key retrieval stay off-limits.

## Safety model

- No instant publishing- publish schedules at least 30 minutes ahead, with a buffer so the window stays honest after the write.
- Every write is reversible: idempotent calls replay stored results on repeat keys, snapshots precede overwrites, and a revert tool exists.
- Every write lands in an audit trail and notifies the token owner in the app.
- Tools outside a token's scope are never registered.
- Every token is bound to an explicit project list.
- Write arguments are allow-listed; safety flags are stamped after the model speaks.

The model is cancel-after, not approve-before. That is adequate for content on your own site that you review; anything landing in someone else's inbox would require a pending-approval flow that has not shipped.