Task/Mcp Usability Improvements

apps/backend/Dockerfile

@@ -40,7 +40,7 @@ ENV UV_COMPILE_BYTECODE=1
 ENV UV_NO_MANAGED_PYTHON=1
 
 # Use copy mode for linking dependencies (required by the cache)
-ENV UV_LINK_MODE=copy 
+ENV UV_LINK_MODE=copy
 # Skip the installation of the dev dependencies
 ENV UV_NO_DEV=1
 
@@ -68,6 +68,8 @@ FROM mirror.gcr.io/library/python:3.10.17-slim AS runtime
 # Copy Bun from official image (lighter than Node.js, required for bunx to run MCP servers)
 COPY --from=oven/bun:latest /usr/local/bin/bun /usr/local/bin/bun
 COPY --from=oven/bun:latest /usr/local/bin/bunx /usr/local/bin/bunx
+# Create npx alias to bunx for compatibility (users can write npx, Docker will use bunx)
+RUN ln -s /usr/local/bin/bunx /usr/local/bin/npx
 
 WORKDIR /app
 

apps/backend/Dockerfile.dev

@@ -51,6 +51,8 @@ COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
 # Copy Bun from official image (lighter than Node.js, required for bunx to run MCP servers)
 COPY --from=oven/bun:latest /usr/local/bin/bun /usr/local/bin/bun
 COPY --from=oven/bun:latest /usr/local/bin/bunx /usr/local/bin/bunx
+# Create npx alias to bunx for compatibility (users can write npx, Docker will use bunx)
+RUN ln -s /usr/local/bin/bunx /usr/local/bin/npx
 
 WORKDIR /app
 

apps/backend/src/rhesis/backend/app/routers/tools.py

@@ -33,23 +33,42 @@ def create_tool(
     current_user: User = Depends(require_current_user_or_token),
 ):
     """
-    Create a new tool integration.
+    Create a new tool.
 
-    The credentials (JSON dict) will be encrypted in the database.
-    Examples: {"NOTION_TOKEN": "ntn_abc..."} or {"GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_abc..."}
+    A tool allows the system to connect to an external service or API. Examples of tools are:
 
-    For custom providers (provider_type="custom"), you must provide the MCP server configuration
-    in tool_metadata with credential placeholders. Placeholders MUST use simple format like
-    {{ TOKEN }} (not {{ TOKEN | tojson }}) because the JSON must be valid before Jinja2 rendering.
+    - MCPs
+    - APIs
+
+    Currently, we support the following MCP tool providers:
+
+    1. **Notion**
+    - Store the Notion token in the credentials dictionary with the key `"NOTION_TOKEN"`.
+    - Example:
+        ```json
+        {"NOTION_TOKEN": "ntn_abc..."}
+        ```
 
-    Example tool_metadata for custom provider:
+    2. **Custom MCP provider**
+    - The MCP server configuration JSON should be provided in `tool_metadata`
+    - The API token should be stored in the credentials dictionary
+    - The custom provider should use **npx** to run the MCP server.
+
+    Example `tool_metadata` for a custom provider:
+    ```json
     {
-        "command": "bunx",
-        "args": ["--bun", "@custom/mcp-server"],
+        "command": "npx",
+        "args": ["@example/mcp-server"],
         "env": {
-            "NOTION_TOKEN": "{{ NOTION_TOKEN }}"
+            "API_TOKEN": "{{ TOKEN }}"
         }
     }
+    ```
+
+    Where the credentials dictionary is:
+    ```json
+    {"TOKEN": "your_api_token_123"}
+    ```
     """
     organization_id, user_id = tenant_context
     return crud.create_tool(db=db, tool=tool, organization_id=organization_id, user_id=user_id)

apps/frontend/src/app/(protected)/mcp/components/MCPConnectionDialog.tsx

@@ -409,7 +409,7 @@ export function MCPConnectionDialog({
                 </Typography>
 
                 <TextField
-                  label="Auth Token"
+                  label="TOKEN"
                   fullWidth
                   required={!isEditMode}
                   type={showAuthToken ? 'text' : 'password'}
@@ -494,9 +494,9 @@ export function MCPConnectionDialog({
                       color="text.secondary"
                       sx={{ mb: 2 }}
                     >
-                      Configure your custom MCP server. Use credential
-                      placeholders with <code>{'{{'}</code> and{' '}
-                      <code>{'}}'}</code> format.
+                      Provide your API token above, then paste your MCP server config below 
+                      using <code>{"{{ TOKEN }}"}</code> as a placeholder 
+                      wherever the token is required.
                     </Typography>
                     <Typography
                       variant="body2"
@@ -519,10 +519,10 @@ export function MCPConnectionDialog({
                       }}
                     >
                       {`{
-  "command": "bunx",
-  "args": ["--bun", "@notionhq/notion-mcp-server"],
+  "command": "npx",
+  "args": ["@example/mcp-server"],
   "env": {
-    "NOTION_TOKEN": "{{ TOKEN }}"
+    "API_TOKEN": "{{ TOKEN }}"
   }
 }`}
                     </Box>

sdk/src/rhesis/sdk/services/mcp/agent.py

@@ -3,8 +3,11 @@
 import asyncio
 import json
 import logging
+from pathlib import Path
 from typing import Any, Dict, List, Optional, Tuple, Union
 
+import jinja2
+
 from rhesis.sdk.models.base import BaseLLM
 from rhesis.sdk.models.factory import get_model
 from rhesis.sdk.services.mcp.client import MCPClient
@@ -28,25 +31,6 @@ class MCPAgent:
     and accomplish tasks. Clients can customize behavior via system prompts.
     """
 
-    DEFAULT_SYSTEM_PROMPT = """You are an autonomous agent that can use MCP tools \
-to accomplish tasks.
-
-You operate in a ReAct loop: Reason → Act → Observe → Repeat
-
-For each iteration:
-- Reason: Think step-by-step about what information you need and how to get it
-- Act: Either call tools to gather information, or finish with your final answer
-- Observe: Examine tool results and plan next steps
-
-Guidelines:
-- Break complex tasks into simple tool calls
-- Use tool results to inform next actions
-- When you have sufficient information, use action="finish" with your final_answer
-- Be efficient: minimize unnecessary tool calls
-- You can call multiple tools in a single iteration if they don't depend on each other
-
-Remember: You must explicitly use action="finish" when done."""
-
     def __init__(
         self,
         model: Optional[Union[str, BaseLLM]] = None,
@@ -69,10 +53,19 @@ def __init__(
         if not mcp_client:
             raise ValueError("mcp_client is required")
 
+        # Initialize template environment
+        templates_dir = Path(__file__).parent / "prompt_templates"
+        self._jinja_env = jinja2.Environment(
+            loader=jinja2.FileSystemLoader(str(templates_dir)),
+            autoescape=False,
+            trim_blocks=True,
+            lstrip_blocks=True,
+        )
+
         # Convert model to BaseLLM instance if needed
         self.model = self._set_model(model)
         self.mcp_client = mcp_client
-        self.system_prompt = system_prompt or self.DEFAULT_SYSTEM_PROMPT
+        self.system_prompt = system_prompt or self._load_default_system_prompt()
         self.max_iterations = max_iterations
         self.verbose = verbose
         self.executor = ToolExecutor(mcp_client)
@@ -83,6 +76,11 @@ def _set_model(self, model: Optional[Union[str, BaseLLM]]) -> BaseLLM:
             return model
         return get_model(model)
 
+    def _load_default_system_prompt(self) -> str:
+        """Load the default system prompt from template."""
+        template = self._jinja_env.get_template("system_prompt.j2")
+        return template.render()
+
     async def run_async(self, user_query: str) -> AgentResult:
         """
         Execute the agent's ReAct loop asynchronously.
@@ -383,34 +381,12 @@ def _build_prompt(
         tools_text = self._format_tools(available_tools)
         history_text = self._format_history(history)
 
-        prompt = f"""User Query: {user_query}
-
-Available MCP Tools:
-{tools_text}
-
-"""
-        if history_text:
-            prompt += f"""Execution History:
-{history_text}
-
-Based on the query, available tools, and execution history above, decide what to do next.
-
-"""
-        else:
-            prompt += """This is the first iteration. Analyze the query and decide \
-what tools to call.
-
-"""
-
-        prompt += """Your response should follow this structure:
-- reasoning: Your step-by-step thinking about what to do
-- action: Either "call_tool" (to execute tools) or "finish" (when you have the answer)
-- tool_calls: List of tools to call if action="call_tool" (can be multiple)
-- final_answer: Your complete answer if action="finish"
-
-Think carefully about what information you need and how to get it efficiently."""
-
-        return prompt
+        template = self._jinja_env.get_template("iteration_prompt.j2")
+        return template.render(
+            user_query=user_query,
+            tools_text=tools_text,
+            history_text=history_text,
+        )
 
     def _format_tools(self, tools: List[Dict[str, Any]]) -> str:
         """Format tool list into human-readable text with names, descriptions, \

sdk/src/rhesis/sdk/services/mcp/client.py

@@ -30,7 +30,7 @@ def __init__(
 
         Args:
             server_name: Friendly name for the server (e.g., "notionApi")
-            command: Command to launch the server (e.g., "bunx", "python")
+            command: Command to launch the server (e.g., "npx", "python")
             args: Command arguments (e.g., ["--bun", "@notionhq/notion-mcp-server"])
             env: Environment variables to pass to the server process
         """
@@ -274,7 +274,7 @@ def from_tool_config(cls, tool_name: str, tool_config: Dict, credentials: Dict[s
 
         Example:
             tool_config = {
-                "command": "bunx",
+                "command": "npx",
                 "args": ["--bun", "@notionhq/notion-mcp-server"],
                 "env": {
                     "NOTION_TOKEN": "{{ NOTION_TOKEN }}"

sdk/src/rhesis/sdk/services/mcp/prompt_templates/iteration_prompt.j2

@@ -0,0 +1,22 @@
+User Query: {{ user_query }}
+
+Available MCP Tools:
+{{ tools_text }}
+
+{% if history_text %}
+Execution History:
+{{ history_text }}
+
+Based on the query, available tools, and execution history above, decide what to do next.
+
+{% else %}
+This is the first iteration. Analyze the query and decide what tools to call.
+
+{% endif %}
+Your response should follow this structure:
+- reasoning: Your step-by-step thinking about what to do
+- action: Either "call_tool" (to execute tools) or "finish" (when you have the answer)
+- tool_calls: List of tools to call if action="call_tool" (can be multiple)
+- final_answer: Your complete answer if action="finish"
+
+Think carefully about what information you need and how to get it efficiently.

sdk/src/rhesis/sdk/services/mcp/prompt_templates/system_prompt.j2

@@ -0,0 +1,17 @@
+You are an autonomous agent that can use MCP tools to accomplish tasks.
+
+You operate in a ReAct loop: Reason → Act → Observe → Repeat
+
+For each iteration:
+- Reason: Think step-by-step about what information you need and how to get it
+- Act: Either call tools to gather information, or finish with your final answer
+- Observe: Examine tool results and plan next steps
+
+Guidelines:
+- Break complex tasks into simple tool calls
+- Use tool results to inform next actions
+- When you have sufficient information, use action="finish" with your final_answer
+- Be efficient: minimize unnecessary tool calls
+- You can call multiple tools in a single iteration if they don't depend on each other
+
+Remember: You must explicitly use action="finish" when done.

sdk/src/rhesis/sdk/services/mcp/provider_templates/atlassian.json.j2

@@ -1,4 +1,4 @@
 {
-  "command": "bunx",
+  "command": "npx",
   "args": ["--bun", "mcp-remote", "https://mcp.atlassian.com/v1/sse"]
 }

sdk/src/rhesis/sdk/services/mcp/provider_templates/notion.json.j2

@@ -1,5 +1,5 @@
 {
-  "command": "bunx",
+  "command": "npx",
   "args": ["--bun", "@notionhq/notion-mcp-server"],
   "env": {
     "NOTION_TOKEN": {{ NOTION_TOKEN | tojson }}