Power Automate MCP

Power Automate via FlowStudio MCP — Foundation

This skill is the plumbing layer. It gives an AI agent a reliable way to talk to a FlowStudio MCP server, discover what tools are available, and handle the responses cleanly. The actual workflow narratives live in four specialized skills that all build on this one.

Real debugging examples: Expression error in child flow | Data entry, not a flow bug | Null value crashes child flow

Requires: A FlowStudio MCP subscription (or compatible Power Automate MCP server). You will need:

• MCP endpoint: https://mcp.flowstudio.app/mcp (same for all subscribers)
• API key / JWT token (x-api-key header — NOT Bearer)
• In ChatGPT or claude.ai there is no key: add https://mcp.flowstudio.app/mcp/oauth as a connector and sign in with Microsoft — see the ChatGPT walkthrough
• Power Platform environment name (e.g. Default-<tenant-guid>)

Which Skill to Use When

Skills are organized by use-case intent, not by which tools they call. Multiple skills reuse the same underlying tools — pick by what the user is trying to accomplish.

Same tools, different lenses. power-automate-build and power-automate-debug both call update_live_flow, get_live_flow, and the run-error tools — they differ in direction (forward vs backward) and intent (compose vs diagnose). power-automate-monitoring and power-automate-governance both call the Store tools — they differ in audience (ops vs compliance) and outcome (read health vs write metadata). Don't try to memorize "which tools belong to which skill"; pick the skill by what the user is doing.

Source of Truth

If documentation disagrees with a real API response, the API wins. Tool schemas in this skill (or any other) may lag the server — call tool_search to confirm the current shape before invoking a tool you haven't used recently.

How Agents Discover Tools

The FlowStudio MCP server (v1.1.5+) exposes two non-billable meta-tools that let an agent load only the tools relevant to the current task. Use these in preference to tools/list (which loads all 30+ schemas at once) or guessing tool names.

The server's tool_search bundles are intentionally narrower than this skill family — they're starter packs of the most-likely-needed tools per intent. A workflow skill (e.g. power-automate-debug) may pull a bundle and then call tool_search again for additional tools as the workflow progresses.

# Cold start — pick a bundle by intent
skills = mcp("list_skills", {})
# [{"name": "debug-flow", "description": "Investigate why a flow is failing...",
#   "tools": ["get_live_flow_runs", "get_live_flow_run_error", ...]}, ...]

# Load schemas for the bundle
debug_tools = mcp("tool_search", {"query": "skill:debug-flow"})

Current common bundles:

Recommended Language: Python or Node.js

All examples in this skill family use Python with urllib.request (stdlib — no pip install needed). Node.js is an equally valid choice: fetch is built-in from Node 18+, JSON handling is native, and async/await maps cleanly onto the request-response pattern of MCP tool calls — making it a natural fit for teams already working in a JavaScript/TypeScript stack.

TL;DR — use the Core MCP Helper (Python or Node.js) below. Both handle JSON-RPC framing, auth, and response parsing in a single reusable function.

Core MCP Helper (Python)

Use this helper throughout all subsequent operations:

import json, urllib.request

TOKEN = "<YOUR_JWT_TOKEN>"
MCP   = "https://mcp.flowstudio.app/mcp"

def mcp(tool, args, cid=1):
    payload = {"jsonrpc": "2.0", "method": "tools/call", "id": cid,
               "params": {"name": tool, "arguments": args}}
    req = urllib.request.Request(MCP, data=json.dumps(payload).encode(),
        headers={"x-api-key": TOKEN, "Content-Type": "application/json",
                 "User-Agent": "FlowStudio-MCP/1.0"})
    try:
        resp = urllib.request.urlopen(req, timeout=120)
    except urllib.error.HTTPError as e:
        body = e.read().decode("utf-8", errors="replace")
        raise RuntimeError(f"MCP HTTP {e.code}: {body[:200]}") from e
    raw = json.loads(resp.read())
    if "error" in raw:
        raise RuntimeError(f"MCP error: {json.dumps(raw['error'])}")
    text = raw["result"]["content"][0]["text"]
    return json.loads(text)

Common auth errors:

• HTTP 401/403 → token is missing, expired, or malformed. Get a fresh JWT from mcp.flowstudio.app.
• HTTP 400 → malformed JSON-RPC payload. Check Content-Type: application/json and body structure.
• MCP error: {"code": -32602, ...} → wrong or missing tool arguments. Call tool_search with select:<toolname> to confirm the schema.

Core MCP Helper (Node.js)

Equivalent helper for Node.js 18+ (built-in fetch — no packages required):

const TOKEN = "<YOUR_JWT_TOKEN>";
const MCP   = "https://mcp.flowstudio.app/mcp";

async function mcp(tool, args, cid = 1) {
  const payload = {
    jsonrpc: "2.0",
    method: "tools/call",
    id: cid,
    params: { name: tool, arguments: args },
  };
  const res = await fetch(MCP, {
    method: "POST",
    headers: {
      "x-api-key": TOKEN,
      "Content-Type": "application/json",
      "User-Agent": "FlowStudio-MCP/1.0",
    },
    body: JSON.stringify(payload),
  });
  if (!res.ok) {
    const body = await res.text();
    throw new Error(`MCP HTTP ${res.status}: ${body.slice(0, 200)}`);
  }
  const raw = await res.json();
  if (raw.error) throw new Error(`MCP error: ${JSON.stringify(raw.error)}`);
  return JSON.parse(raw.result.content[0].text);
}

Requires Node.js 18+. For older Node, replace fetch with https.request from the stdlib or install node-fetch.

Verify the Connection

A 3-line smoke test that confirms the token, endpoint, and helper all work:

skills = mcp("list_skills", {})
print(f"Connected — {len(skills)} skill bundles available:",
      [s["name"] for s in skills])

Expected output:

Connected — 6 skill bundles available: ['build-flow', 'create-flow', 'debug-flow', 'monitor-flow', 'discover', 'governance']

If this fails, see the Common auth errors note above. If it succeeds, hand off to the workflow skill matching the user's intent.

Handling Oversized Responses

Some MCP tool responses are large enough to overflow the agent's context window:

When the harness spills to a file

Agent harnesses (Claude Code, VS Code Copilot, etc.) save oversized responses to a temp file (e.g. tool-results/mcp-flowstudio-describe_live_connector-NNNN.txt) and return the path instead of the inline JSON. The file is double-wrapped — the outer MCP envelope plus the inner JSON-escaped payload:

[{"type":"text","text":"<JSON-escaped payload>"}]

Two parses to reach a usable object:

import json
with open(path) as f:
    raw = json.loads(f.read())
payload = json.loads(raw[0]["text"])

$payload = ((Get-Content $path -Raw | ConvertFrom-Json)[0].text) | ConvertFrom-Json

Rules of thumb

1. Extract, don't echo. Pull the specific field(s) you need (one operationId, one action's outputs) and discard the rest before reasoning about it.
2. Always pass actionName to get_live_flow_run_action_outputs. Omitting it fetches all top-level actions. For actions inside a foreach, passing actionName without iterationIndex can return every repetition of that action.
3. Reuse the spill file within a session. Refetching the same connector swagger costs 30+ seconds and produces another spill — cache the path.
4. Don't grep the spill file for JSON keys directly. Strings are JSON-escaped inside the file (\"OperationId\":), so a plain grep for "OperationId": will not match. Parse first, then filter.
5. Summarize tool output to the user. Echo name + state + trigger for flow lists and actionName + status + code for run errors — not raw JSON, unless asked.

# Good — drill into one operation in a connector swagger
conn = mcp("describe_live_connector", {"environmentName": ENV, "connectorName": "shared_sharepointonline"})
op = conn["properties"]["swagger"]["paths"]["/datasets/{dataset}/tables/{table}/items"]["get"]
print(op["operationId"], "—", op.get("summary"))

# Bad — keeping the whole 500 KB swagger in context
print(json.dumps(conn, indent=2))   # don't do this

Auth & Connection Notes

Reference Files

• MCP-BOOTSTRAP.md — endpoint, auth, request/response format (read this first)
• tool-reference.md — response shapes and behavioral notes (parameters are in tool_search)
• action-types.md — Power Automate action type patterns
• connection-references.md — connector reference guide