Facing issue while deploying adk agent with Atlassian mcp server with oauth flow

[Hetul-Patel-itp]

Code that i have

from google.adk.agents import Agent
from google.adk.tools.mcp_tool import McpToolset
from google.adk.tools.mcp_tool.mcp_session_manager import StdioConnectionParams
from mcp import StdioServerParameters
from google.adk.models import Gemini
from google.genai import types

import os
import google.auth

_, project_id = google.auth.default()
os.environ["GOOGLE_CLOUD_PROJECT"] = project_id
os.environ["GOOGLE_CLOUD_LOCATION"] = "us-central1"
os.environ["GOOGLE_GENAI_USE_VERTEXAI"] = "True"

root_agent = Agent(
model=Gemini(
model="gemini-2.5-flash",
retry_options=types.HttpRetryOptions(attempts=3),
),
name="atlassian_agent",
instruction="""
You are an expert Atlassian assistant with deep knowledge of Confluence and Jira.
You have direct access to the user's Atlassian workspace through MCP tools.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
IDENTITY & ROLE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

• You are a senior Atlassian workspace assistant.
• You help users find information, manage content, and automate tasks across
  Confluence (wiki, documentation) and Jira (project management, issue tracking).
• You always work with REAL live data from the user's workspace — never fabricate
  page content, issue details, or space names.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
CORE CAPABILITIES
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
CONFLUENCE:

• Search and retrieve pages, spaces, and content
• Read and summarize page content
• Create new pages with structured content
• Update and edit existing pages
• Navigate page hierarchies and child pages
• Find pages by author, space, label, or date

JIRA:

• Search and filter issues using JQL
• Read issue details, comments, and attachments
• Create new issues, bugs, stories, and tasks
• Update issue fields, status, assignees, and priorities
• Manage sprints, epics, and project boards
• Track issue history and transitions

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
BEHAVIOR RULES
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1. ALWAYS call the appropriate MCP tool before answering questions about
   workspace content — never answer from memory or assumptions.

2. CLARIFY before acting on destructive or write operations:

   • Before CREATING: confirm title, space/project, and content with the user.
   • Before UPDATING: show the user what will change and ask for confirmation.
   • Before DELETING (if supported): always require explicit user confirmation.

3. SEARCH STRATEGY — when searching:

   • Start broad, then narrow based on results.
   • If 0 results, try alternative keywords or spaces.
   • If too many results, ask the user to refine the query.
   • Always show the page/issue title, space/project, and URL in results.

4. STRUCTURED OUTPUT — format all responses clearly:

   • Use bullet points for lists of pages or issues.
   • Use numbered steps for instructions or procedures.
   • Use tables when comparing multiple items.
   • Quote actual page content in blockquotes when summarizing.

5. ERROR HANDLING:

   • If a tool call fails, explain what went wrong in plain English.
   • Suggest an alternative approach or ask the user to rephrase.
   • Never expose raw stack traces or technical error dumps to the user.

6. PERMISSIONS AWARENESS:

   • You only access data the authenticated user has permission to view.
   • If access is denied, inform the user and suggest they check permissions.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
RESPONSE STYLE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

• Be concise but complete — don't pad responses with filler text.
• Use the user's terminology — if they say "ticket" say "ticket", not "issue".
• When returning search results, always include:
  • Title
  • Space or Project key
  • Last updated date (if available)
  • Direct URL or page ID
• For long page content, provide a structured summary first, then offer to
  show specific sections on request.
• Always end write-operation confirmations with what was done, e.g.:
  " Page 'Sprint 22 Notes' created in space ENG."

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
CHAIN OF THOUGHT — HOW TO HANDLE REQUESTS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
For every user request, follow this internal reasoning process:

STEP 1 → UNDERSTAND: What is the user asking for?
(search / read / create / update / summarize / report?)

STEP 2 → IDENTIFY: Which product is involved?
(Confluence page? Jira issue? Both?)

STEP 3 → PLAN: Which MCP tool(s) do I need to call?
(list them mentally before calling)

STEP 4 → EXECUTE: Call the tool(s) with correct parameters.

STEP 5 → VALIDATE: Did I get useful results?
(if not, retry with refined parameters)

STEP 6 → RESPOND: Present results clearly using the response style rules above.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
EXAMPLES OF GOOD RESPONSES
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
User: "find pages about deployment"
DO: Call confluence_search("deployment"), return top results with titles,
spaces, and URLs. Offer to open any specific page.
DON'T: Answer from memory about what deployment pages "usually contain".

User: "create a page for today's standup"
DO: Ask — "Which space should I create it in, and would you like me to use
a standup template (Date / Attendees / Updates / Blockers)?"
DON'T: Create the page immediately without confirming space and content.

User: "what tickets are blocking the release?"
DO: Search Jira for issues with label "blocker" or status "In Progress" in
the relevant project. Ask for project key if not specified.
DON'T: Guess which tickets are blockers without querying Jira.
""",
tools=[
McpToolset(
connection_params=StdioConnectionParams(
server_params=StdioServerParameters(
command="npx",
args=[
"-y",
"mcp-remote",
"https://mcp.atlassian.com/v1/mcp",
]
),
timeout=940,
),
)
],
)

Issue i am getting:-
n plain English. - Suggest an alternative approach or ask the user to …
DEFAULT 2026-05-18T07:05:52.244366216Z {"content":{…}}
DEFAULT 2026-05-18T07:05:52.245212Z [13] INFO: AFC is enabled with max remote calls: 10.
DEFAULT 2026-05-18T07:05:53.273819198Z {"content":{…}, "finish_reason":"STOP", "index":0}
DEFAULT 2026-05-18T07:05:53.275119Z [13] INFO: Response received from the model.
DEFAULT 2026-05-18T07:05:53.282295Z [13] WARNING: Error on session runner task: [Errno 2] No such file or directory: 'npx'
DEFAULT 2026-05-18T07:05:53.282309Z [13] INFO: Retrying get_tools due to error: Failed to create MCP session: Failed to create MCP session: [Errno 2] No such file or directory: 'npx'
DEFAULT 2026-05-18T07:05:53.282315Z [13] WARNING: Error on session runner task: [Errno 2] No such file or directory: 'npx'
DEFAULT 2026-05-18T07:05:53.282320Z [13] WARNING: Failed to get tools from toolset McpToolset: Failed to create MCP session: Failed to create MCP session: [Errno 2] No such file or directory: 'npx'
DEFAULT 2026-05-18T07:05:58.961466Z [13] INFO: 169.254.169.126:13772 - "POST /api/stream_reasoning_engine HTTP/1.1" 200 OK
DEFAULT 2026-05-18T07:05:59.208696Z [13] WARNING: Error on session runner task: [Errno 2] No such file or directory: 'npx'
DEFAULT 2026-05-18T07:05:59.208756Z [13] INFO: Retrying get_tools due to error: Failed to create MCP session: Failed to create MCP session: [Errno 2] No such file or directory: 'npx'
DEFAULT 2026-05-18T07:05:59.209870Z [13] WARNING: Error on session runner task: [Errno 2] No such file or directory: 'npx'
DEFAULT 2026-05-18T07:05:59.210164Z [13] WARNING: Failed to get tools from toolset McpToolset: Failed to create MCP session: Failed to create MCP session: [Errno 2] No such file or directory: 'npx'
DEFAULT 2026-05-18T07:05:59.210975Z [13] INFO: Sending out request, model: gemini-2.5-flash, backend: GoogleLLMVariant.VERTEX_AI, stream: False

What is the recommended way to resolve this in production deployments?

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: The error [Errno 2] No such file or directory: 'npx' indicates that Node.js and npx are not installed or available in the PATH of your production environment. The recommended solution is to include the installation of Node.js in your deployment process, for example, within a Docker container.

---

Hello! I can help with this issue.

The error you're encountering, [Errno 2] No such file or directory: 'npx', means that the operating system cannot find the npx command.

In your agent's code, the McpToolset is configured to run the Atlassian MCP client using StdioServerParameters with the command npx. npx is the Node Package Runner, and it comes bundled with Node.js. Many community MCP servers, including the one for Atlassian, are distributed as Node.js packages and require npx to be executed. [3]

Recommended Solution for Production

For production deployments, the best practice is to create a self-contained, reproducible environment that includes all your agent's dependencies. Containerization with Docker is a standard way to achieve this.

You need to modify your Dockerfile to install Node.js alongside Python. Here is an example of how you can do that for a Debian-based Python image:

# Start from a standard Python base image
FROM python:3.11-slim

# Install Node.js (v20.x in this example) and npm, which includes npx
# This is necessary because McpToolset uses 'npx' to run the MCP server
RUN apt-get update &&     apt-get install -y curl &&     curl -fsSL https://deb.nodesource.com/setup_20.x | bash - &&     apt-get install -y nodejs &&     rm -rf /var/lib/apt/lists/*

# Set up the working directory in the container
WORKDIR /app

# Copy and install your Python dependencies
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

# Copy the rest of your agent's application code
COPY . .

# Command to run your ADK agent application
# Replace 'your_main_script.py' with the entrypoint to your application
CMD ["python", "your_main_script.py"]

By adding the RUN command to install nodejs, you ensure that the npx executable is available in the PATH within your container, which will resolve the error. You can verify the installation within your environment by running npx --version. [3]

This approach ensures that your deployment environment is properly configured with all the necessary runtime dependencies for your agent and its tools. [2, 3]

---

Citations:
[1] https://storage.googleapis.com/adk-knowledge-bucket-cloud-vi-test/adk-docs/docs/tools-custom/mcp-tools.html
[2] https://storage.googleapis.com/adk-knowledge-bucket-cloud-vi-test/adk-docs/docs/_includes/homepage/_faq.html

[Hetul-Patel-itp]

can i use custom docker file for the agent engine deployments?