UiPath Orchestrator MCP服务器

uipath-orchestrator-mcp-python

Python MCP server for UiPath Orchestrator — drive your automations from Claude Desktop and any MCP-aware agent.

What is this?

A Model Context Protocol server that exposes your UiPath Orchestrator as a set of agent tools. Point Claude Desktop (or Cline, Continue, or a custom agent) at it and ask, in plain language, to start jobs, inspect failures, manage queues, and read audit logs — across Cloud, on-prem, and PAT deployments.

It's a typed async Python client (25 tools, Pydantic models, three production auth strategies) wrapped as a native MCP server — not a REST wrapper, not a bespoke integration.

Why does it exist?

REST APIs aren't MCP servers — an agent can't discover or call them without glue code. And the MCP servers that do exist don't cover UiPath Orchestrator, especially on-prem and sovereign deployments where most examples stop at SaaS. This fills that gap: production auth coverage (Cloud OAuth · On-prem · PAT) behind a single MCP surface.

What you can ask the agent

Once it's connected, these are the kinds of prompts you can send:

• "Schedule the Invoice Processing job to run weekdays at 9 AM in the Finance folder."
• "List all jobs that failed in the last 24 hours, grouped by process, with the error message for the most common failure."
• "How many items are pending in the OnboardingEmployees queue, and which robots are configured to process it?"
• "Run a one-time execution of the Reconciliation process in the AsiaTenant, then notify me when it finishes."
• "Pull yesterday's audit log entries for any folder permission changes."
• "Show me the assets defined in the HR folder that look like credentials (don't print their values)."
• "Give me a health snapshot of the Production folder — job states, queue backlogs, robot availability."

Install (Claude Desktop)

1. Clone and install:

git clone https://github.com/Mo3bdlaa/uipath-orchestrator-mcp-python.git
cd uipath-orchestrator-mcp-python
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -e .

2. Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "uipath-orchestrator": {
      "command": "python",
      "args": ["-m", "uipath_mcp_python.server"],
      "env": {
        "UIPATH_URL": "https://cloud.uipath.com/your-org/your-tenant/",
        "UIPATH_AUTH_TYPE": "cloud-oauth",
        "UIPATH_CLIENT_ID": "...",
        "UIPATH_CLIENT_SECRET": "...",
        "UIPATH_TENANT_NAME": "Default"
      }
    }
  }
}

Use the absolute path to your virtualenv's interpreter for command (e.g. /path/to/uipath-orchestrator-mcp-python/.venv/bin/python) so Claude Desktop launches the server with the dependencies installed. After pip install -e . you can also use the uipath-mcp console script instead of python -m uipath_mcp_python.server.

Config file location:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Restart Claude Desktop and the uipath-orchestrator tools appear in the 🔌 menu.

Configure auth

Pick the one strategy that matches your deployment. All strategies also need UIPATH_URL and UIPATH_TENANT_NAME.

On-prem is the deployment most MCP/UiPath examples never cover — and it's the one that matters for regulated and sovereign environments. It's a first-class strategy here, not an afterthought.

For local development, copy .env.example → .env and fill in the values instead of using the env block above.

Tools exposed

The server exposes 25 tools, grouped by Orchestrator domain:

Folders & Infrastructure

Assets

Processes & Jobs

Queues

Schedules

Logs & Audit

Analytics

The descriptions above mirror the tool docstrings in uipath_mcp_python/server.py — that file is the source of truth as tools change.

Tools are annotated with MCP hints (readOnlyHint / destructiveHint) so an agent host can tell safe reads from state changes. Only start_process, enqueue_item, and cancel_job mutate Orchestrator state.

Uniform list responses. Every list/query tool returns the same envelope, so an agent never has to guess the shape:

{ "items": [ /* rows */ ], "count": 2, "total": 57 }

count is the size of the returned page; total is the server-side total when the OData endpoint reports one (otherwise null).

Resources (read-only endpoints)

Parameterised resource templates

Reliability & transport

The client is built for unattended agent use:

• Token refresh is serialised behind a lock and reused with a 5-minute buffer, so concurrent tool calls trigger a single handshake. A 401 triggers one transparent refresh-and-retry.
• Transient failures — both HTTP 429/5xx responses and transport-level errors (connection resets, read/connect timeouts) — are retried with exponential backoff and jitter, honouring the Retry-After header when present.
• Timeouts and a bounded connection pool are configured on the HTTP client; error bodies are trimmed before surfacing as a typed OrchestratorError.
• Observability: retries, token refreshes, and give-ups are logged through the uipath_mcp_python logger (silent by default — attach a handler / set a level in your host app to see them).

Optional tuning via environment variables:

Project structure

uipath_mcp_python/
├── schemas.py        # Pydantic models for every Orchestrator entity
├── orchestrator.py   # Async API client with 3 auth strategies
├── settings.py       # Environment-to-config loader
└── server.py         # MCP tool & resource definitions

See examples/ for worked agent interactions, including a multi-step list-then-schedule flow and an on-prem deployment.

Development

pip install -e ".[dev]"
pytest                              # unit tests — no live Orchestrator required
ruff check uipath_mcp_python/ tests/
mypy                                # static type check (config in pyproject.toml)

The package ships a py.typed marker, so downstream projects get full type information when importing uipath_mcp_python.

License

MIT — free to use, modify, and distribute.

Built by Mohammed Shaker