Skip to content

ravuthz/export-post-mcp

v0.1.0

Export accessible AI conversations and related assets as portable Markdown post ZIPs.

GitHub

Export Post MCP

An MCP service that packages a conversation supplied by an AI agent into a portable post-{slug}.zip. The ZIP contains post.md, README.md, an optional chronological handoff, and supplied assets. It follows the format in the two Export Post ZIPs provided for this project.

An MCP server cannot discover the host application's current conversation or attachments on its own. The calling agent must send the accessible user/assistant messages, write the finished post Markdown, and supply files as base64 or explicitly allowed local paths. This design works across MCP-capable agents without claiming access to other chats or hidden instructions.

Run locally

Requires Python 3.10 or newer and no third-party packages.

python3 server.py

The process speaks MCP over standard input/output. Logs and errors go to standard error. The included mcp.json configures the same command for plugin hosts that support local stdio MCP servers. Set EXPORT_POST_OUTPUT_DIR for ZIP output (default: the output directory beside server.py). To allow local_path assets, set EXPORT_POST_ALLOWED_ROOTS to a list of absolute directories separated by the OS path separator. No local file path is allowed by default.

Run over HTTP

python3 server.py --http --host 127.0.0.1 --port 8765

This provides a stateless MCP endpoint at http://127.0.0.1:8765/mcp. For ChatGPT on the web, connect it through a Secure MCP Tunnel or deploy it behind HTTPS. Do not expose an unprotected local test server to the public internet. Binding outside loopback requires EXPORT_POST_HTTP_TOKEN; clients must send Authorization: Bearer <token>. Set EXPORT_POST_PUBLIC_BASE_URL to the public HTTPS origin when you want tool results to include a signed download URL for the ZIP. The download link grants access to that ZIP to anyone holding it.

The HTTP mode supports JSON responses to MCP POST requests. It does not implement OAuth or server push. A client that requires those features needs an adapter or a later server extension.

Example client configuration:

{
  "mcpServers": {
    "export-post": {
      "command": "python3",
      "args": ["/absolute/path/to/export-post-mcp/server.py"],
      "env": {
        "EXPORT_POST_OUTPUT_DIR": "/absolute/path/to/exports",
        "EXPORT_POST_ALLOWED_ROOTS": "/absolute/path/to/approved/assets"
      }
    }
  }
}

Replace the paths before adding this config to an MCP client.

Install and connect (export-post-mcp.appkh.online)

The MCP service is hosted and exposed over HTTPS via Cloudflare Tunnel at export-post-mcp.appkh.online.

  • Endpoint URL: https://export-post-mcp.appkh.online/mcp
  • Transport: MCP JSON-RPC 2.0 over HTTP POST (stateless)
  • Public Downloads: Completed exports return signed download URLs under https://export-post-mcp.appkh.online/downloads/{export_id}/{token}

1. Claude Code CLI

Add the MCP server directly using the CLI:

claude mcp add export-post https://export-post-mcp.appkh.online/mcp

2. Cursor, Windsurf, or Direct HTTP Clients

Add the remote endpoint to your client configuration (e.g. ~/.cursor/mcp.json or project .cursor/mcp.json):

{
  "mcpServers": {
    "export-post": {
      "url": "https://export-post-mcp.appkh.online/mcp"
    }
  }
}

3. Claude Desktop (macOS / Windows)

For clients that use local stdio processes, bridge to the remote HTTP endpoint using mcp-remote in your claude_desktop_config.json:

{
  "mcpServers": {
    "export-post": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://export-post-mcp.appkh.online/mcp"]
    }
  }
}

4. Verify Connection

Verify that the remote endpoint is responsive and serving MCP tools:

# Initialize handshake
curl -s -X POST https://export-post-mcp.appkh.online/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc": "2.0", "id": 1, "method": "initialize", "params": {"protocolVersion": "2024-11-05", "capabilities": {}, "clientInfo": {"name": "test-client", "version": "1.0.0"}}}'

# List available tools
curl -s -X POST https://export-post-mcp.appkh.online/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc": "2.0", "id": 2, "method": "tools/list", "params": {}}'

Process management (PM2)

Manage background services (export-post-mcp and export-post-tunnel) configured in ecosystem.config.cjs using the Makefile:

# Start or restart services
make start
make restart

# Inspect status and metadata
make status
make info

# Stream logs (or tail without streaming)
make logs
make logs-tail LINES=100
make logs-mcp
make logs-tunnel

# Stop or remove services from PM2
make stop
make delete

# Persist process list across reboots
make save
make resurrect

# Interactive monitor
make monit

Run make help for the full list of available targets and overrides.

Tools

  • export_post: one call for a title, finished Markdown, accessible messages, and small inline assets.
  • begin_export: start a staged export and receive an export_id.
  • append_messages: send accessible user and assistant messages in chronological order.
  • add_asset: send a base64 file or a path inside EXPORT_POST_ALLOWED_ROOTS.
  • finalize_export: supply or update post.md, validate relative links, and create the ZIP.
  • get_export: inspect a draft or retrieve the finished ZIP path.

The service never accepts system/developer/tool/private-reasoning messages. It rejects obvious credentials and private keys, unsafe paths, missing local Markdown targets, oversized files, and malformed base64. Review post.md and the ZIP before publishing; automated secret checks cannot recognize every secret.

Try it

make test
make package

Alternatively, run the underlying commands directly:

python3 -m unittest discover -s tests -v
python3 package_plugin.py

The package command creates export-post-mcp-plugin.zip with the local MCP runtime and skill. Copy or install that package on each machine that should run the service. ChatGPT on the web requires a remotely reachable MCP endpoint or a Secure MCP Tunnel; a local stdio process alone is not available to a web chat.

The final response from a calling agent should link to the returned ZIP path, report the main file post.md, count the packaged files, and mention any expected assets it could not access.