Connect AI Tools (MCP)

Core has a built-in MCP server, so AI tools and bots — Claude Code, Claude Desktop, Cursor, or your own agent — can find your workflows, run them, and draft new ones. The Desktop assistant uses this same server.

Connection details

Address	http://127.0.0.1:4002/mcp (change the port with TIPELINE_MCP_PORT)
Transport	Streamable HTTP
Auth	Authorization: Bearer <token>
Token	Core's API token, saved in ~/Tipeline/.database/core.token

Core must be running: the Desktop app starts it, or run tipeline start.

Claude Code

claude mcp add --transport http tipeline http://127.0.0.1:4002/mcp \
  --header "Authorization: Bearer $(cat ~/Tipeline/.database/core.token)"

Cursor and other clients with HTTP support

{
  "mcpServers": {
    "tipeline": {
      "url": "http://127.0.0.1:4002/mcp",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

Clients that only start local servers (Claude Desktop)

Bridge to the HTTP server with mcp-remote:

{
  "mcpServers": {
    "tipeline": {
      "command": "npx",
      "args": ["mcp-remote", "http://127.0.0.1:4002/mcp",
               "--header", "Authorization: Bearer <token>"]
    }
  }
}

Tools

Tool	What it does
search_workflows	Finds workflows for a request — installed ones, and ones on your enabled hubs that aren't installed yet — with their public and protected inputs and, when they mark them, their outputs
list_tasks	Every installed workflow with its public and protected inputs. Drafts are marked draft: true
exec_task	Runs an installed workflow by identifier with optional variable values, waits for it, and returns { status, pipelineIdentifier, durationMs, outputs, outputDetails?, description?, error? }
authoring_guide	How to write a workflow in YAML, with examples
list_actions	Every action group, built-in and plugin, with what each action does
describe_actions	The fields, return type and an example of named actions, e.g. ["http.request"]
check_workflow	Checks workflow YAML the way installing and running it would; returns { valid, errors, warnings }
create_draft_workflow	Saves workflow YAML as a draft after checking it

A typical bot asks search_workflows with the user's request, fills in the inputs, and calls exec_task. To build something new: authoring_guide → search_workflows (to reuse installed workflows as sub-workflow steps) → list_actions → describe_actions → check_workflow → create_draft_workflow, then a person tests the draft and marks it ready. check_workflow also checks sub-workflow steps against the workflows they call.

What a client can and can't do

• Runs show up everywhere. A run started over MCP appears live in the Desktop app and the CLI, and in the workflow's run history, like any other run.
• Callers see public and protected. exec_task returns the final values of the workflow's non-secret outputs, public or protected — or, when it marks none, of all its public, non-secret variables. outputDetails gives each output's label and description (what the value means, such as its unit) and description the workflow's own, so a client can explain the result. Search results list public and protected inputs, plus the outputs a workflow marks. Private and secret variables are never shared.
• Drafts never run. exec_task refuses a draft, and search leaves drafts out. A person has to test it and click Mark as ready.
• Hub workflows aren't installed by the client. A search result with a hub field isn't installed yet; someone installs it (the Desktop assistant offers to) before it can run.
• Approvals still apply. A workflow with an approval step waits for a person to decide.

Security

The token is full access

Core's API token is the same token its WebSocket port accepts. Anyone holding it can run any installed workflow on that Core — with any variable values, including private ones (the same as --var in the CLI) — and save drafts. Give it only to tools you trust, and keep core.token private.

• The MCP server only accepts connections from this machine. To reach it from others, set TIPELINE_BIND_HOST before Core starts, and block port 4002 with a firewall from anyone who doesn't need it.
• For several people, or when you need roles and an audit trail, put the Tipeline Server in front of Core instead of sharing the Core token.
• If port 4002 is taken, Core starts without the MCP server and logs a warning; set TIPELINE_MCP_PORT to another port, and point your clients (and the assistant's MCP connection setting) at it.