# FlowStudio MCP — Power Automate Skills for AI Agents Give your AI agent the same visibility you have in the Power Automate portal — plus a bit more. The Graph API only returns top-level run status — agents can't see action inputs, loop iterations, or nested failures. Flow Studio MCP exposes all of it. ![Agent debugging a Power Automate flow via MCP](assets/demo-debug.gif) **You can click through the portal and find the root cause. Your agent can't — unless it has MCP.** ![The portal shows everything to a human — but agents only see the top-level error via Graph API](assets/portal-vs-reality.png) ![With Flow Studio MCP, the agent sees what you see](assets/mcp-root-cause.png) ## When you need this - Your agent can see that a flow failed, but not why — Graph API only returns status codes - You want your agent to see action-level inputs and outputs, like you can in the portal - A loop has hundreds of iterations and some produced bad output — in the portal you'd click through each one, but the agent can scan all iteration inputs and outputs at once - You want to check flow health, failure rates, and maker activity across your tenant without opening the admin center - You need to classify flows, detect orphaned resources, or audit connectors at scale — without installing the CoE Starter Kit - You're tired of being the middle-man between your agent and the portal ## Graph API vs Flow Studio MCP The core difference: **Graph API gives your agent run status. MCP gives your agent the inputs and outputs of every action.** | What the agent sees | Graph API | Flow Studio MCP | | --------------------------------------------- | --------- | --------------- | | Run passed or failed | Yes | Yes | | **Action inputs and outputs** | **No** | **Yes** | | Error details beyond status code | No | Yes | | Child flow run details | No | Yes | | Loop iteration data | No | Yes | | Flow definition (read + write) | Limited | Full JSON | | Resubmit / cancel runs | Limited | Yes | | **Cached flow health & failure rates** | **No** | **Yes** | | **Maker / Power Apps / connection inventory** | **No** | **Yes** | | **Governance metadata (tags, impact, owner)** | **No** | **Yes** | ## Skills | Skill | Description | | ---------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- | | [`power-automate-mcp`](skills/power-automate-mcp/) | Connect to and operate Power Automate cloud flows — list flows, read definitions, check runs, resubmit, cancel | | [`power-automate-debug`](skills/power-automate-debug/) | Step-by-step diagnostic process for investigating failing flows | | [`power-automate-build`](skills/power-automate-build/) | Build, scaffold, and deploy Power Automate flow definitions from scratch | | [`power-automate-monitoring`](skills/power-automate-monitoring/) | Flow health, failure rates, maker inventory, Power Apps, environment and connection counts | | [`power-automate-governance`](skills/power-automate-governance/) | Classify flows by impact, detect orphans, audit connectors, manage notifications, compute archive scores | The first three skills use **live** Power Automate API calls. The monitoring and governance skills use the **cached store** — a daily snapshot with aggregated stats, remediation hints, and governance metadata. Requires a FlowStudio for Teams or MCP Pro+ subscription for store tools. Each skill follows the [Agent Skills specification](https://agentskills.io/specification) and works with any compatible agent. ### Supported agents Copilot, Claude Code, Codex, OpenClaw, Gemini CLI, Cursor, Goose, Amp, OpenHands ## Quick Start ### Install as Claude Code plugin Available through the Claude plugin marketplace after approval. To test locally: ```bash git clone https://github.com/ninihen1/power-automate-mcp-skills.git claude --plugin-dir ./power-automate-mcp-skills ``` Then connect the MCP server: ```bash claude mcp add --transport http flowstudio https://mcp.flowstudio.app/mcp \ --header "x-api-key: " ``` Get your token at [mcp.flowstudio.app](https://mcp.flowstudio.app). ### Install in Codex Inside a Codex session, install skills directly: ``` $skill-installer install https://github.com/ninihen1/power-automate-mcp-skills/tree/main/skills/power-automate-mcp $skill-installer install https://github.com/ninihen1/power-automate-mcp-skills/tree/main/skills/power-automate-debug $skill-installer install https://github.com/ninihen1/power-automate-mcp-skills/tree/main/skills/power-automate-build $skill-installer install https://github.com/ninihen1/power-automate-mcp-skills/tree/main/skills/power-automate-monitoring $skill-installer install https://github.com/ninihen1/power-automate-mcp-skills/tree/main/skills/power-automate-governance ``` Then connect the MCP server in `~/.codex/config.toml`: ```toml [mcp_servers.flowstudio] url = "https://mcp.flowstudio.app/mcp" [mcp_servers.flowstudio.http_headers] x-api-key = "" ``` ### Install via skills.sh Search for [flowstudio on skills.sh](https://skills.sh/?q=flowstudio), or: ```bash npx skills add github/awesome-copilot -s flowstudio-power-automate-mcp npx skills add github/awesome-copilot -s flowstudio-power-automate-debug npx skills add github/awesome-copilot -s flowstudio-power-automate-build npx skills add github/awesome-copilot -s flowstudio-power-automate-monitoring npx skills add github/awesome-copilot -s flowstudio-power-automate-governance ``` ### Install via ClawHub ```bash npx clawhub@latest install power-automate-mcp ``` ### Install via Smithery ```bash npx smithery skill add flowstudio/power-automate-mcp ``` ### Manual install Copy the skill folder(s) into your project's `.github/skills/` directory (or wherever your agent discovers skills). ### Connect the MCP server **Claude Code:** ```bash claude mcp add --transport http flowstudio https://mcp.flowstudio.app/mcp \ --header "x-api-key: " ``` **Codex** (`~/.codex/config.toml`): ```toml [mcp_servers.flowstudio] url = "https://mcp.flowstudio.app/mcp" [mcp_servers.flowstudio.http_headers] x-api-key = "" ``` **Copilot / VS Code** (`.vscode/mcp.json`): ```json { "servers": { "flowstudio": { "type": "http", "url": "https://mcp.flowstudio.app/mcp", "headers": { "x-api-key": "" } } } } ``` Get your token at [mcp.flowstudio.app](https://mcp.flowstudio.app). ## Real debugging examples These are from real production investigations, not demos. - **[Expression error in child flow](examples/fix-expression-error.md)** — `contains(string(...))` crashed on a nested property. Agent traced through parent flow, into child, through loop iterations, and found the failing input. Portal showed "ExpressionEvaluationFailed" with no context. - **[Data entry, not a flow bug](examples/data-not-flow.md)** — User reported two "bugs" back to back. Agent proved both were data entry errors (missing comma in email, single address in CC field). Flow was correct. Diagnosed in seconds. - **[Null value crashes child flow](examples/null-child-flow.md)** — `split(Name, ', ')` crashed when 38% of records had null Names. Agent traced parent to child to loop to action, found the root cause, and deployed a fix via `update_live_flow`. ## Prerequisites - A [FlowStudio](https://mcp.flowstudio.app) MCP subscription (all live tools) - For store tools (monitoring, governance): FlowStudio for Teams or MCP Pro+ - MCP endpoint: `https://mcp.flowstudio.app/mcp` - API key / JWT token (passed as `x-api-key` header) ## Repository structure ``` skills/ power-automate-mcp/ core connection & operation skill power-automate-debug/ debug workflow skill power-automate-build/ build & deploy skill power-automate-monitoring/ flow health & tenant inventory skill power-automate-governance/ compliance & governance skill examples/ real debugging walkthroughs README.md LICENSE MIT ``` ## Available on GitHub Works with Copilot, Claude, and any MCP-compatible agent. - [awesome-copilot](https://github.com/github/awesome-copilot) (merged) - [skills.sh](https://skills.sh/?q=flowstudio) (3K+ installs) - [Smithery](https://smithery.ai/skills/flowstudio/power-automate-mcp) (published) - [ClawHub](https://clawhub.ai) (v1.1.0) ## Contributing Contributions welcome. Each skill folder must contain a `SKILL.md` with the required frontmatter. See the existing skills for the format. ## License [MIT](LICENSE) --- Keywords: Power Automate debugging, flow run history, expression evaluation failed, child flow failure, nested action errors, loop iteration output, agent automation MCP, Power Platform AI, flow definition deploy, resubmit failed run, flow monitoring, governance, CoE, orphan detection, connector audit, archive score, maker inventory