flowstudio-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.

| The user wants to… | Load this skill |

|---|---|

| Make or change a flow (build new, modify existing, fix a bug, deploy) | flowstudio-power-automate-build |

| Diagnose why a flow failed (root cause analysis on a failing run) | flowstudio-power-automate-debug |

| See tenant-wide flow health, failure rates, asset inventory | flowstudio-power-automate-monitoring (Pro+) |

| Tag, audit, classify, score, or offboard flows | flowstudio-power-automate-governance (Pro+) |

| Just connect, set up auth, write the helper, parse responses | this skill (foundation) |

Same tools, different lenses. flowstudio-power-automate-build and flowstudio-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).

flowstudio-power-automate-monitoring and flowstudio-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

| Priority | Source | Covers |

|----------|--------|--------|

| 1 | Real API response | Always trust what the server actually returns |

| 2 | tool_search / list_skills | Authoritative tool schemas, parameter names, types, required flags |

| 3 | SKILL docs & reference files | Workflow narrative, response shapes, non-obvious behaviors |

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.

| Meta-tool | When to call |

|---|---|

| list_skills | Cold start — see the available bundles (build-flow, create-flow, debug-flow, monitor-flow, discover, governance) and pick one |

| tool_search with query: "skill:<name>" | Load the full schema set for one bundle (e.g. skill:debug-flow) |

| tool_search with query: "select:tool1,tool2" | Load specific tools by name (e.g. when chaining across bundles) |

| tool_search with query: "<keywords>" | Free-text search when the user request is ambiguous (e.g. "cancel run") |

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. flowstudio-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:

| Bundle | Use when |

|---|---|

| create-flow | Creating a brand-new flow; includes environment/connection discovery, connector description, dynamic options, and update_live_flow |

| build-flow | Reading or modifying an existing flow definition |

| debug-flow | Investigating failed runs and action-level inputs/outputs |

| monitor-flow | Starting/stopping, triggering, cancelling, or resubmitting runs |

| discover | Enumerating environments, flows, and connections |

| governance | Pro+ cached-store tagging, maker audit, and metadata updates |

---

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.

| Language | Verdict | Notes |

|---|---|---|

| Python | Recommended | Clean JSON handling, no escaping issues, all skill examples use it |

| Node.js (≥ 18) | Recommended | Native fetch + JSON.stringify/JSON.parse; no extra packages |

| PowerShell | Avoid for flow operations | ConvertTo-Json -Depth silently truncates nested definitions; quoting and escaping break complex payloads. Acceptable for a quick connectivity smoke-test but not for building or updating flows. |

| cURL / Bash | Possible but fragile | Shell-escaping nested JSON is error-prone; no native JSON parser |

> 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:

| Tool | Typical size | Cause |

|---|---|---|

| describe_live_connector | 100-600 KB | Full Swagger spec for a connector |

| get_live_dynamic_properties | 50-500 KB | Dynamic connector field schemas such as SharePoint list columns |

| get_live_flow_run_action_outputs (no actionName) | 50 KB – several MB | Top-level action outputs; with an action in a foreach, every repetition can be returned |

| get_live_flow (large flows) | 50-500 KB | Deeply nested branches |

| list_live_flows (large tenants) | 50-200 KB | Hundreds of flow records |

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

| Field | Value |

|---|---|

| Auth header | x-api-key: <JWT> — not Authorization: Bearer |

| Token format | Plain JWT — do not strip, alter, or prefix it |

| Timeout | Use ≥ 120 s for get_live_flow_run_action_outputs (large outputs) |

| Environment name | Default-<tenant-guid> (find it via list_live_environments or list_live_flows response) |

---

Reference Files

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