Pathista

What is an MCP workflow? - Pathista help center

Help pages: What is an MCP workflow?

Updated

An MCP workflow is a process written as steps that an AI assistant works through by calling an MCP server. The server keeps the position and hands over only the current step, so during a run the assistant sees no later step. Pathista is such a server: you write the workflow once and run it in ChatGPT or Claude.

What MCP is

MCP, the Model Context Protocol, is an open-source standard for connecting AI applications to external systems. An assistant that supports it, such as ChatGPT or Claude, can call the tools an MCP server offers.

MCP defines how the calls are made. What a server does with them is up to the server. A workflow server uses them to keep a process in order.

How a run goes

The position, the variables, the notes and the history stay in your Pathista account, not in the chat, so a run can be continued in a later message or a later chat.

  1. You write a workflow once: steps, the ways from one step to the next (loops are allowed) and, for each step, an instruction and, if you want one, a checklist.
  2. In your assistant you ask it to start the workflow by name. Pathista starts a run and hands back the first step.
  3. The assistant does the work of the step and closes each checklist item with evidence of what it did.
  4. It moves on. Where a step has several ways out, the assistant names the one it takes, and a condition on a way can keep it closed until a variable allows it.
  5. The final step is closed like any other, and only then is the run done.

What the server adds to a plain prompt

Pathista checks that evidence is given and is not a generic word; it does not judge whether the evidence is true. What the assistant wrote stays in the run's log for you to read.

  • One step at a time: during a run the assistant is handed the current step and the ways out of it, not a later step.
  • Checklists that need evidence: a todo is closed only with evidence of at least 15 characters, and a bare “done” is refused.
  • Bounded loops: a way back can carry a condition such as iteration < maxIterations (a counter of rounds and the limit the template sets for it), so a loop ends.
  • A record: every move, note and closed todo is in the run's log, which you read on the run's page.

Where it runs

ChatGPT (as a custom MCP server), Claude (as a custom connector) and Claude Code are the tested assistants. Connect an assistant has the address and the steps; it also names other MCP clients that can use the same address.

Pathista runs the workflows in your account: the ones you wrote, imported or added from a template.

A real workflow

9 steps

The Release checklist for a Claude Code project template has 9 steps and 22 checklist items. Its steps, in the order they are written:

Words in code type, such as testsGreen, are the template's own variables: small yes/no values or counters that the assistant sets during the run and that decide which way out of a step is open.

  1. What ships
  2. Changelog, version and release commit
  3. Build and tests on the release commit
  4. Fix the failing tests
  5. Tag the release commit
  6. Confirm the publish with the user
  7. Publish
  8. Verify the release from a clean setup
  9. Stop the release and report

The checklist of the step “Build and tests on the release commit”. The assistant may leave the step only when every item is closed with evidence:

  • The command, the commit id it ran on and the passed, failed and skipped counts are quoted from the output
  • Every failing test is named with the first line of its error, or the words no failures
  • The context variable testsGreen is set to the result (true or false)

Open the template

Rejoining the server...

Rejoin failed... trying again in seconds.

Failed to rejoin.
Please retry or reload the page.

The session has been paused by the server.

Failed to resume the session.
Please retry or reload the page.