Connecting MCP

The MCP server is an HTTP endpoint authenticated with a Bearer token. Through it an agent reads and changes nodes directly, without leaving its workflow: tools like create_node, get_node and rag_query show up in its tool list.

The server address is the same for everyone: https://mcp.flow.shipoora.ru/mcp.

Step 1. Get a token

1. Sign in to the web app.
2. Open Settings → MCP and click Generate access token, giving it a name — say, my-laptop. The name is what lets you revoke this one token later instead of all of them.
3. Copy the token immediately. It looks like aiflow_pat_... and is shown once — only its hash is stored on the server, so there is nothing to show again.

The same page carries ready-made configs with the token already filled in, each with a copy button.

Step 2. Connect a client

Everything at once — the npm package

npx @flownix/cli mcp --token aiflow_pat_...

The package verifies the token against the server before writing any config: a client configured with a bad token costs half an hour of searching for nothing. With no arguments (npx @flownix/cli) it also installs the skills and the reporter hook described further down this page.

It writes config only for the clients whose format is documented here — Claude Code and Cursor; for the rest it prints a ready snippet to paste.

Claude Code — one command

claude mcp add --transport http flownix https://mcp.flow.shipoora.ru/mcp \
  --header "Authorization: Bearer aiflow_pat_..." \
  --scope user

--scope user makes the server available in all of your projects. The alternatives are local (the default, current directory only) and project — don't use that one with a token: project writes to .mcp.json, a file meant to be committed, and your token rides along into the repository.

Verify — the server should answer:

claude mcp list

In a session the tools appear as mcp__flownix__create_node, mcp__flownix__rag_query and so on. The prefix is whatever name you registered the server under (flownix above). The server never checks its own name, so one connected under the old ai-flow key keeps working — renaming it is optional.

Claude Code — by config

If you edit the config by hand, the fragment looks like this:

{
  "mcpServers": {
    "flownix": {
      "url": "https://mcp.flow.shipoora.ru/mcp",
      "headers": { "Authorization": "Bearer aiflow_pat_..." }
    }
  }
}

Cursor

Same shape; the file is ~/.cursor/mcp.json for every project, or .cursor/mcp.json inside one:

{
  "mcpServers": {
    "flownix": {
      "url": "https://mcp.flow.shipoora.ru/mcp",
      "headers": { "Authorization": "Bearer aiflow_pat_..." }
    }
  }
}

Any other client

The transport is Streamable HTTP and authentication is the Authorization: Bearer <token> header. Nothing in the protocol is client-specific.

Step 3. Confirm the server answers

When the client stays silent and it isn't clear whose problem it is, ask the server directly:

curl -X POST https://mcp.flow.shipoora.ru/mcp \
  -H "Authorization: Bearer aiflow_pat_..." \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize",
       "params":{"protocolVersion":"2025-11-25",
       "capabilities":{},"clientInfo":{"name":"my-agent","version":"1.0"}}}'

A response carrying result.serverInfo means the token was accepted and the server works — what's left to debug is the client config. A 401 means the token is wrong or revoked.

What's next

A connected server is access, nothing more. For the agent to work by the rules rather than improvise, it needs skills: installing skills. After that — your first project.

Token reporter

The agent presence panel shows who is working on what. How many tokens were spent is known only to the client — the MCP server cannot see it by design — so a separate script reports it. The script is optional: without it presence works in full, one field stays empty, and the panel states why it is empty instead of drawing a zero.

npx @flownix/cli hook install --token <your-token>

curl -fsSL https://flow.shipoora.ru/flownix-reporter.sh | bash -s -- install <your-token>

The installer places the script in ~/.claude (or the directory from CLAUDE_CONFIG_DIR), registers it as Stop and SessionEnd hooks in settings.json, and keeps a copy of your previous settings next to it. It leaves the status line alone: if another plugin configured one, it keeps working.

npx @flownix/cli hook check