Instruction file imported from lucis/gistify (.cursor/rules/deco-chat.mdc). Copyright stays with the author.

deco.chat Platform Development Rules

🎯 Overview

deco.chat is a powerful platform for creating MCP (Model Context Protocol) tools and workflows that can automate any kind of workload. This rule file provides comprehensive guidance for developing applications on the deco.chat platform.

🚀 Getting Started

Prerequisites

• Deno installed on your system
• Access to deco.chat platform
• Basic TypeScript knowledge

Initial Setup Commands

# Install CLI
deno run -A jsr:@deco/cli

# Configure project
deno run -A jsr:@deco/cli config

# Development
npm run dev

# Deploy
npm run deploy

Project Structure

project/
├── main.ts           # Main application
├── deco.gen.ts       # Auto-generated types
├── package.json      # Dependencies
├── wrangler.toml     # Cloudflare config
└── README.md

🔧 Core Concepts

1. Environment (Env) Object

The env object contains all available integrations and tools. Each integration is namespaced:

interface Env {
  INTEGRATION_NAME: {
    TOOL_NAME: (input: ToolInput) => Promise<ToolOutput>;
  };
  DECO_CHAT_WORKSPACE_API: {
    // Workspace-specific tools
  };
  DECO_CHAT_API: {
    // Global API tools
  };
}

2. Tools

Tools are individual functions that perform specific tasks. They can:

• Call external APIs
• Process data
• Interact with databases
• Generate content using AI

3. Workflows

Workflows orchestrate multiple tools using Mastra's control flow patterns. They should:

• Split I/O operations into separate steps
• Use control flow for data processing and logic
• Keep each step focused on a single tool invocation

📝 Creating Tools

Core Imports

import { withRuntime } from "@deco/workers-runtime";
import {
  createStepFromTool,
  createTool,
  createWorkflow,
} from "@deco/workers-runtime/mastra";
import { z } from "zod";
import { Env } from "./deco.gen";

Basic Tool Pattern

const createMyTool = (env: Env) =>
  createTool({
    id: "TOOL_ID",
    description: "Tool description",
    inputSchema: z.object({
      param1: z.string(),
      param2: z.number().optional(),
    }),
    outputSchema: z.object({
      result: z.string(),
    }),
    execute: async ({ context }) => {
      // ONE tool call only
      const response = await env.INTEGRATION.TOOL({
        // parameters
      });

      return { result: response.data };
    },
  });

Tool Best Practices

1. Single Responsibility: Each tool should do one thing well
2. Input Validation: Use Zod schemas to validate inputs
3. Error Handling: Always handle potential errors gracefully
4. Type Safety: Leverage TypeScript for type safety

🔄 Creating Workflows

Workflow Philosophy - GOLDEN RULE

IMPORTANT: Follow this principle for optimal workflow design:

• Each step should invoke only ONE tool from env.INTEGRATION_NAME.TOOL_NAME
• All data processing, logic, and control flow should be handled by Mastra workflow operators
• Split complex operations into multiple steps rather than combining them

Basic Workflow Pattern

const createMyWorkflow = (env: Env) => {
  const step1 = createStepFromTool(createTool1(env));
  const step2 = createStepFromTool(createTool2(env));

  return createWorkflow({
    id: "WORKFLOW_ID",
    inputSchema: z.object({ input: z.string() }),
    outputSchema: z.object({ output: z.string() }),
  })
    .then(step1)
    .map((context) => ({
      ...context,
      // Data processing here
    }))
    .then(step2)
    .commit();
};

Control Flow Operators

1. Sequential (.then)

.then(step1)
.then(step2)
.then(step3)

2. Parallel (.parallel)

.parallel([
  step1,
  step2,
  step3
])

3. Conditional (.branch)

.branch(
  (context) => condition,
  (workflow) => workflow.then(stepIfTrue),
  (workflow) => workflow.then(stepIfFalse)
)

4. Data Transformation (.map)

.map((context) => ({
  ...context,
  processedData: context.rawData.toUpperCase(),
  isValid: context.rawData.length > 0
}))

5. Looping (.repeat)

.repeat(
  (context) => context.items,
  processItem
)

🔁 Looping with .dountil

What is .dountil?

The .dountil operator allows you to repeat a workflow step (typically a polling or status-checking tool) until a specified condition is met. This is especially useful for workflows that need to wait for an asynchronous task to complete, such as polling an external API or a long-running browser automation.

How does it work?

• You provide a step (created with createStepFromTool) that will be executed repeatedly.
• You provide a predicate function that receives the output of the repeated step (as inputData) and returns true when the loop should stop.

Usage Pattern

.then(scrapeStep)
.dountil(
  pollTaskStep,
  async ({ inputData }) => {
    // inputData is the output of pollTaskStep
    return inputData.status === "finished";
  }
)
.map(async ({ getStepResult }) => ({
  // Use getStepResult(pollTaskStep) to access the final output
  output: getStepResult(pollTaskStep).output,
  prompt: getStepResult(pollTaskStep).prompt,
  taskId: getStepResult(pollTaskStep).taskId,
}))

Key Points

• The first argument to .dountil is the step to repeat.
• The second argument is an async predicate function that receives an object with inputData, which is the output of the repeated step.
• The loop continues until the predicate returns true.
• After .dountil, you can use .map and getStepResult(pollTaskStep) to access the final output.

Real-World Example

Suppose you want to scrape a website, poll for the scraping task to finish, and then process the result:

const pollTaskStep = createStepFromTool(pollTaskTool);

return createWorkflow({
  // ...
})
  .then(scrapeStep)
  .dountil(
    pollTaskStep,
    async ({ inputData }) => {
      // inputData is the output of pollTaskStep
      return inputData.status === "finished";
    }
  )
  .map(async ({ getStepResult }) => ({
    // getStepResult(pollTaskStep) gives you the final poll result
    output: getStepResult(pollTaskStep).output,
    prompt: getStepResult(pollTaskStep).prompt,
    taskId: getStepResult(pollTaskStep).taskId,
  }))
  // ... continue workflow

Best Practices

• Use .dountil for polling or waiting scenarios.
• Always check the output (inputData) of your polling step for the completion condition.
• Use .map and getStepResult after .dountil to extract the final result for downstream steps.

Summary Table for .dountil:

In short:
.dountil(step, predicate) repeats step until predicate({ inputData }) returns true. Afterward, use .map and getStepResult(step) to access the final output.

🗺️ Data Transformation with .map and Accessing Previous Step Results

The .map operator allows you to transform, enrich, or combine data between tool steps. Use .map to:

• Prepare or reformat data for the next tool
• Combine results from multiple previous steps
• Add custom logic or computed fields

Basic Usage of .map

.then(fetchUserData)
.map(({ inputData }) => ({
  // Add a computed field
  ...inputData,
  isActive: inputData.status === "active",
}))
.then(processUser)

Accessing Previous Step Results with getStepResult

When you need to access the output of a specific previous step (not just the immediate last one), use the getStepResult function provided in the .map context.

const stepA = createStepFromTool(toolA(env));
const stepB = createStepFromTool(toolB(env));

return createWorkflow({
  // ...
})
  .then(stepA)
  .then(stepB)
  .map(({ inputData, getStepResult }) => {
    const resultA = getStepResult(stepA);
    const resultB = getStepResult(stepB);
    return {
      combined: `${resultA.value} + ${resultB.value}`,
    };
  })

• getStepResult(step) retrieves the output of any previous step.
• This is especially useful when you need to merge or reference data from multiple steps.

Real-World Example

In a workflow that summarizes a YouTube video, you might want to combine the video search result, transcript, and summary into a single output. Here’s how you can do it:

const searchStep = createStepFromTool(createYouTubeSearchTool(env));
const transcriptStep = createStepFromTool(createYouTubeTranscriptTool(env));
const summarizeStep = createStepFromTool(createSummarizeTextTool(env));

return createWorkflow({
  id: "YOUTUBE_VIDEO_SUMMARY",
  inputSchema: z.object({ query: z.string() }),
  outputSchema: z.object({
    videoId: z.string(),
    title: z.string(),
    description: z.string(),
    transcript: z.string(),
    summary: z.string(),
  }),
})
  .then(searchStep)
  .map(async ({ inputData }) => ({ videoId: inputData.videoId }))
  .then(transcriptStep)
  .map(async ({ inputData, getStepResult }) => {
    const searchResult = getStepResult(searchStep);
    return {
      videoId: searchResult.videoId,
      title: searchResult.title,
      description: searchResult.description,
      transcript: inputData.transcript,
    };
  })
  .then(summarizeStep)
  .map(async ({ inputData, getStepResult }) => {
    const searchResult = getStepResult(searchStep);
    const transcriptResult = getStepResult(transcriptStep);
    return {
      videoId: searchResult.videoId,
      title: searchResult.title,
      description: searchResult.description,
      transcript: transcriptResult.transcript,
      summary: inputData.summary,
    };
  })
  .commit();

Best Practices

• Use .map for all data processing, not inside tool execute functions.
• Use getStepResult to keep your workflow logic clear and maintainable.
• Keep each step focused on a single responsibility and use .map to orchestrate data flow between steps.

🔌 Integration Usage Patterns

AI Generation

const aiResponse = await env.GIMENES.AGENT_GENERATE_TEXT({
  message: "Your prompt",
  options: { model: "gpt-4" }
});

Database Operations

const dbResult = await env.DECO_CHAT_WORKSPACE_API.DATABASES_RUN_SQL({
  sql: "SELECT * FROM table",
  params: [param1, param2]
});

File System Operations

const fileContent = await env.DECO_CHAT_WORKSPACE_API.FS_READ({
  path: "/path/to/file",
  expiresIn: 3600
});

📋 Configuration Files

wrangler.toml

main = "main.ts"
compatibility_date = "2025-06-17"
compatibility_flags = [ "nodejs_compat" ]

[deco]
app = "your-app"
workspace = "your-workspace"
enable_workflows = true

[[deco.bindings]]
name = "INTEGRATION_NAME"
type = "mcp"
integration_id = "your-id"

[[migrations]]
tag = "v1"
new_classes = [ "Workflow" ]

[durable_objects]
[[durable_objects.bindings]]
name = "DECO_CHAT_WORKFLOW_DO"
class_name = "Workflow"

package.json

{
  "scripts": {
    "dev": "deco dev",
    "deploy": "wrangler deploy --dry-run --outdir dist && cd dist && deco deploy"
  },
  "dependencies": {
    "@deco/workers-runtime": "npm:@jsr/deco__workers-runtime@^0.2.20",
    "zod": "^3.24.2"
  },
  "devDependencies": {
    "@deco/cli": "npm:@jsr/deco__cli@^0.5.12"
  }
}

🎯 Golden Rules

1. One Tool Per Step: Each workflow step should call only ONE env.INTEGRATION.TOOL()
2. Control Flow for Logic: Use .map(), .branch(), .parallel() for data processing
3. Split Complex Operations: Break down into multiple focused steps
4. Validate Inputs: Always use Zod schemas
5. Handle Errors: Use branching for error scenarios
6. Type Safety: Leverage TypeScript and generated types

🔍 Common Patterns

Data Validation

.map((context) => ({
  ...context,
  isValid: context.data && context.data.length > 0,
  dataCount: context.data?.length || 0
}))

Error Handling

.branch(
  (context) => context.isValid,
  (workflow) => workflow.then(processData),
  (workflow) => workflow.map((context) => ({
    ...context,
    success: false,
    error: "Invalid data"
  }))
)

Parallel Processing

.parallel([
  fetchUserProfile,
  fetchUserPreferences,
  fetchUserActivity
])
.map((context) => ({
  ...context,
  hasCompleteData: context.profile && context.preferences && context.activity
}))

Conditional Processing

.branch(
  (context) => context.dataType === "users",
  (workflow) => workflow.then(processUsers),
  (workflow) => workflow.then(processProducts)
)

🚨 Anti-Patterns to Avoid

❌ Multiple Tool Calls in One Step

// DON'T DO THIS
execute: async ({ context }) => {
  const data1 = await env.API1.TOOL1({...});
  const data2 = await env.API2.TOOL2({...});
  const data3 = await env.API3.TOOL3({...});
  return { result: "combined" };
}

❌ Complex Logic in Tools

// DON'T DO THIS
execute: async ({ context }) => {
  const data = await env.API.TOOL({...});

  // Complex processing logic here
  const processed = data.map(item => ({
    ...item,
    calculated: item.value * 2,
    formatted: item.name.toUpperCase(),
    validated: item.id > 0
  })).filter(item => item.validated);

  return { result: processed };
}

✅ Do This Instead

// Step 1: Fetch data
const fetchData = createStepFromTool(createFetchTool(env));

// Step 2: Process data (in workflow)
.then(fetchData)
.map((context) => ({
  ...context,
  processed: context.data.map(item => ({
    ...item,
    calculated: item.value * 2,
    formatted: item.name.toUpperCase(),
    validated: item.id > 0
  })).filter(item => item.validated)
}))

📚 Practical Examples

Example 1: User Data Processing Pipeline

❌ Bad Approach (Don't Do This)

// DON'T: Combining multiple tool calls in one step
const createBadUserProcessor = (env: Env) =>
  createTool({
    id: "BAD_USER_PROCESSOR",
    description: "Process user data - BAD EXAMPLE",
    inputSchema: z.object({ userId: z.string() }),
    outputSchema: z.object({ result: z.string() }),
    execute: async ({ context }) => {
      // Multiple tool calls in one step - AVOID THIS
      const userData = await env.DECO_CHAT_WORKSPACE_API.DATABASES_RUN_SQL({
        sql: "SELECT * FROM users WHERE id = ?",
        params: [context.userId]
      });

      const aiAnalysis = await env.GIMENES.AGENT_GENERATE_TEXT({
        message: `Analyze user data: ${JSON.stringify(userData)}`
      });

      const savedResult = await env.DECO_CHAT_WORKSPACE_API.FS_WRITE({
        path: `/users/${context.userId}/analysis.json`,
        content: aiAnalysis.text,
        contentType: "application/json"
      });

      return { result: "processed" };
    },
  });

✅ Good Approach (Do This)

// DO: Split into separate steps, use control flow for processing
const createUserDataWorkflow = (env: Env) => {
  // Step 1: Fetch user data (ONE tool call)
  const fetchUserData = createStepFromTool(
    createTool({
      id: "FETCH_USER_DATA",
      description: "Fetch user data from database",
      inputSchema: z.object({ userId: z.string() }),
      outputSchema: z.object({ userData: z.any() }),
      execute: async ({ context }) => {
        const result = await env.DECO_CHAT_WORKSPACE_API.DATABASES_RUN_SQL({
          sql: "SELECT * FROM users WHERE id = ?",
          params: [context.userId]
        });
        return { userData: result };
      },
    })(env)
  );

  // Step 2: Analyze with AI (ONE tool call)
  const analyzeUserData = createStepFromTool(
    createTool({
      id: "ANALYZE_USER_DATA",
      description: "Analyze user data with AI",
      inputSchema: z.object({ userData: z.any() }),
      outputSchema: z.object({ analysis: z.string() }),
      execute: async ({ context }) => {
        const result = await env.GIMENES.AGENT_GENERATE_TEXT({
          message: `Analyze this user data: ${JSON.stringify(context.userData)}`
        });
        return { analysis: result.text || "" };
      },
    })(env)
  );

  // Step 3: Save results (ONE tool call)
  const saveAnalysis = createStepFromTool(
    createTool({
      id: "SAVE_ANALYSIS",
      description: "Save analysis to file system",
      inputSchema: z.object({
        userId: z.string(),
        analysis: z.string()
      }),
      outputSchema: z.object({ saved: z.boolean() }),
      execute: async ({ context }) => {
        await env.DECO_CHAT_WORKSPACE_API.FS_WRITE({
          path: `/users/${context.userId}/analysis.json`,
          content: context.analysis,
          contentType: "application/json"
        });
        return { saved: true };
      },
    })(env)
  );

  return createWorkflow({
    id: "USER_DATA_WORKFLOW",
    inputSchema: z.object({ userId: z.string() }),
    outputSchema: z.object({
      success: z.boolean(),
      analysis: z.string(),
      saved: z.boolean()
    }),
  })
    .then(fetchUserData)
    .map((context) => ({
      ...context,
      // Data processing logic here (not in tools)
      processedUserData: context.userData.results?.[0] || null,
      hasValidData: context.userData.results?.length > 0
    }))
    .branch(
      (context) => context.hasValidData,
      (workflow) => workflow
        .then(analyzeUserData)
        .map((context) => ({
          ...context,
          // More data processing
          analysisLength: context.analysis.length,
          isLongAnalysis: context.analysis.length > 500
        }))
        .then(saveAnalysis)
        .map((context) => ({
          ...context,
          success: true
        })),
      (workflow) => workflow
        .map((context) => ({
          ...context,
          success: false,
          analysis: "No user data found",
          saved: false
        }))
    )
    .commit();
};

Example 2: Multi-Service Integration

const createMultiServiceWorkflow = (env: Env) => {
  // Step 1: Fetch data from external API
  const fetchExternalData = createStepFromTool(
    createTool({
      id: "FETCH_EXTERNAL_DATA",
      description: "Fetch data from external service",
      inputSchema: z.object({ query: z.string() }),
      outputSchema: z.object({ externalData: z.any() }),
      execute: async ({ context }) => {
        const result = await env.EXTERNAL_API.GET_DATA({
          query: context.query
        });
        return { externalData: result };
      },
    })(env)
  );

  // Step 2: Process with AI
  const processWithAI = createStepFromTool(
    createTool({
      id: "PROCESS_WITH_AI",
      description: "Process data with AI",
      inputSchema: z.object({ data: z.any() }),
      outputSchema: z.object({ processedData: z.string() }),
      execute: async ({ context }) => {
        const result = await env.GIMENES.AGENT_GENERATE_TEXT({
          message: `Process this data: ${JSON.stringify(context.data)}`
        });
        return { processedData: result.text || "" };
      },
    })(env)
  );

  return createWorkflow({
    id: "MULTI_SERVICE_WORKFLOW",
    inputSchema: z.object({
      query: z.string(),
      userId: z.string()
    }),
    outputSchema: z.object({
      success: z.boolean(),
      processedData: z.string()
    }),
  })
    .then(fetchExternalData)
    .map((context) => ({
      ...context,
      // Data validation and processing
      isValidData: context.externalData && context.externalData.length > 0,
      dataCount: context.externalData?.length || 0
    }))
    .branch(
      (context) => context.isValidData,
      (workflow) => workflow
        .then(processWithAI)
        .map((context) => ({
          ...context,
          success: true
        })),
      (workflow) => workflow
        .map((context) => ({
          ...context,
          success: false,
          processedData: "No valid data found"
        }))
    )
    .commit();
};

Example 3: Parallel Processing

const createParallelProcessingWorkflow = (env: Env) => {
  // Step 1: Fetch user profile
  const fetchUserProfile = createStepFromTool(
    createTool({
      id: "FETCH_USER_PROFILE",
      description: "Fetch user profile data",
      inputSchema: z.object({ userId: z.string() }),
      outputSchema: z.object({ profile: z.any() }),
      execute: async ({ context }) => {
        const result = await env.DECO_CHAT_WORKSPACE_API.DATABASES_RUN_SQL({
          sql: "SELECT * FROM user_profiles WHERE user_id = ?",
          params: [context.userId]
        });
        return { profile: result.results?.[0] };
      },
    })(env)
  );

  // Step 2: Fetch user preferences
  const fetchUserPreferences = createStepFromTool(
    createTool({
      id: "FETCH_USER_PREFERENCES",
      description: "Fetch user preferences",
      inputSchema: z.object({ userId: z.string() }),
      outputSchema: z.object({ preferences: z.any() }),
      execute: async ({ context }) => {
        const result = await env.DECO_CHAT_WORKSPACE_API.DATABASES_RUN_SQL({
          sql: "SELECT * FROM user_preferences WHERE user_id = ?",
          params: [context.userId]
        });
        return { preferences: result.results?.[0] };
      },
    })(env)
  );

  return createWorkflow({
    id: "PARALLEL_PROCESSING_WORKFLOW",
    inputSchema: z.object({ userId: z.string() }),
    outputSchema: z.object({
      success: z.boolean(),
      dataSummary: z.any()
    }),
  })
    .parallel([
      fetchUserProfile,
      fetchUserPreferences
    ])
    .map((context) => ({
      ...context,
      // Combine and validate parallel results
      hasCompleteData: context.profile && context.preferences,
      dataSummary: {
        profileComplete: !!context.profile,
        preferencesComplete: !!context.preferences
      }
    }))
    .branch(
      (context) => context.hasCompleteData,
      (workflow) => workflow
        .map((context) => ({
          ...context,
          success: true
        })),
      (workflow) => workflow
        .map((context) => ({
          ...context,
          success: false
        }))
    )
    .commit();
};

🔍 Debugging and Testing

Local Development

# Start with debugging
deco dev --debug

# Check logs
deco logs

# Test specific workflow
deco test workflow-name

Common Issues

1. Type Errors: Check your Zod schemas and TypeScript types
2. Integration Errors: Verify integration configuration and credentials
3. Workflow Errors: Check step dependencies and data flow
4. Deployment Errors: Verify wrangler.toml configuration

📚 Best Practices Summary

1. Workflow Design

• Keep steps atomic: Each step should do one thing
• Use control flow for logic: Leverage .map, .branch, .parallel
• Separate concerns: I/O operations vs data processing
• Handle errors gracefully: Use try-catch and proper error schemas

2. Tool Design

• Validate inputs: Always use Zod schemas
• Provide clear descriptions: Help users understand what your tool does
• Return consistent outputs: Use predictable data structures
• Handle edge cases: Consider what happens when things go wrong

3. Integration Usage

• Check documentation: Each integration has specific parameters
• Use TypeScript: Leverage the generated types for type safety
• Handle rate limits: Be mindful of API limitations
• Cache when appropriate: Avoid unnecessary repeated calls

4. Performance

• Parallelize when possible: Use .parallel for independent operations
• Batch operations: Group related operations together
• Optimize data flow: Minimize data transformation overhead

📖 Additional Resources

• Mastra Workflows: https://mastra.ai/en/docs/workflows/control-flow
• Data Mapping: https://mastra.ai/en/docs/workflows/input-data-mapping
• deco.chat Documentation: Platform-specific guides and examples
• TypeScript Documentation: For advanced type usage
• Zod Documentation: For schema validation patterns

🎯 Key Takeaways

Remember the core principle: Each workflow step should invoke only ONE tool, and all other processing should use Mastra's control flow operators. This approach ensures:

• Maintainability: Clear separation of concerns
• Testability: Each step can be tested independently
• Scalability: Easy to modify and extend workflows
• Performance: Optimal execution patterns

Start simple, build incrementally, and leverage the power of deco.chat's integration ecosystem to create powerful automation workflows!

• Scalability: Easy to modify and extend workflows
• Performance: Optimal execution patterns

Start simple, build incrementally, and leverage the power of deco.chat's integration ecosystem to create powerful automation workflows!