
# Runtime Variables

Runtime variables are used to manage private information, so you never write sensitive data directly into tool code or expose it to the LLM.

## What Are Runtime Variables

When a tool needs sensitive information (such as API keys, account passwords, etc.), prefer runtime variables. **Never type sensitive information directly into a conversation with the AI.**

You can also use runtime variables as a general configuration feature, making it easy to share settings across tools.

## Usage Guide

### 1. Configure Variables

Open the app and go to **Settings > Runtime Variables**.

![Runtime variables](/images/jelly-studio/runtime_vars.png)

On this page you can:
- Add new variables.
- Edit existing variables.
- Delete variables you no longer use.

### 2. Use Them in a Conversation

When talking with the LLM, simply tell the AI to use a certain variable. The tool automatically searches for and checks whether that variable exists.

```txt
Help me build an email-sending tool that supports a subject line, body editing, and attachment upload.

Note: get the email account, password, and related parameters from the `email` variable. Do not ask the user to enter them.
```

![User variables in chat](/images/jelly-studio/user_vars_in_chat.png)

### 3. Reference Variables

When you reference a variable in a conversation, the AI automatically recognizes it and reads the corresponding value—no need to type sensitive information manually.

> **Security note**: runtime variables are stored locally. They are never uploaded to the cloud or sent to the LLM, which better protects your information.


Runtime Variables Runtime Variables Runtime variables are used to manage private information, so you never write sensitive data directly into tool code or expose it to the LLM. What Are Runtime Variables When a tool needs sensitive information (such as API keys, account passwords, etc.), prefer runtime variables. Never type sensitive information directly into a conversation with the AI. You can also use runtime variables as a general configuration feature, making it easy to share settings across tools. Usage Guide Configure Variables Open the app and go to Settings > Runtime Variables . Runtime variables On this page you can: Add new variables. Edit existing variables. Delete variables you no longer use. Use Them in a Conversation When talking with the LLM, simply tell the AI to use a certain variable. The tool automatically searches for and checks whether that variable exists. Help me build an email-sending tool that supports a subject line, body editing, and attachment upload. Note: get the email account, password, and related parameters from the email variable. Do not ask the user to enter them. User variables in chat Reference Variables When you reference a variable in a conversation, the AI automatically recognizes it and reads the corresponding value—no need to type sensitive information manually. Security note : runtime variables are stored locally. They are never uploaded to the cloud or sent to the LLM, which better protects your information. Runtime Variables What Are Runtime Variables Usage Guide 1. Configure Variables 2. Use Them in a Conversation 3. Reference Variables

Runtime Variables

Runtime Variables What Are Runtime Variables When a tool needs sensitive information (such as API keys, account passwords, etc.), prefer runtime variables. Never type sensitive information directly into a conversation with the AI. You can also use runtime variables as a general configuration feature, making it easy to share settings across tools.

What Are Runtime Variables

Usage Guide

Runtime Variables 1. Configure Variables Open the app and go to Settings > Runtime Variables . Runtime variables On this page you can: Add new variables. Edit existing variables. Delete variables you no longer use.

1. Configure Variables

Runtime Variables 2. Use Them in a Conversation When talking with the LLM, simply tell the AI to use a certain variable. The tool automatically searches for and checks whether that variable exists. Help me build an email-sending tool that supports a subject line, body editing, and attachment upload. Note: get the email account, password, and related parameters from the email variable. Do not ask the user to enter them. User variables in chat

2. Use Them in a Conversation

Runtime Variables 3. Reference Variables When you reference a variable in a conversation, the AI automatically recognizes it and reads the corresponding value—no need to type sensitive information manually. Security note : runtime variables are stored locally. They are never uploaded to the cloud or sent to the LLM, which better protects your information.

3. Reference Variables

examples/email_send Email Tool and Variables Email-sending tools usually involve sensitive information. Do not provide the email account, password, and similar details directly in a conversation with the AI. Instead, create the relevant variables in runtime variables first and configure the required parameters there. For variable configuration, see Runtime Variables. Text Examples Use an AI Conversation to Sort Out the Variable Fields Needed for Sending Email I want to build an email-sending tool. What information does an email usually need to be configured with? Which parts are suitable to store in variables, and which are better entered as parameters at runtime? For the ones stored as variables, how should the fields typically be named—especially the English naming? Use Variables to Generate the Email-Sending Tool Use the actual key you configured in "Runtime Variables" as the variable name. For example, you can use email, or a naming style that suits you better such as my company.email. Help me build a small email-sending tool. It should support entering a subject, editing the body, and uploading attachments. Note: get the email account, password, and related information from the email variable. Do not ask the user to enter them manually. Email Tool and Variables Text Examples Use an AI Conversation to Sort Out the Variable Fields Needed for Sending Email Use Variables to Generate the Email-Sending Tool

examples/email_send

examples/email_send Email Tool and Variables Email-sending tools usually involve sensitive information. Do not provide the email account, password, and similar details directly in a conversation with the AI. Instead, create the relevant variables in runtime variables first and configure the required parameters there. For variable configuration, see Runtime Variables.

Email Tool and Variables

Text Examples

examples/email_send Use an AI Conversation to Sort Out the Variable Fields Needed for Sending Email I want to build an email-sending tool. What information does an email usually need to be configured with? Which parts are suitable to store in variables, and which are better entered as parameters at runtime? For the ones stored as variables, how should the fields typically be named—especially the English naming?

Use an AI Conversation to Sort Out the Variable Fields Needed for Sending Email

examples/email_send Use Variables to Generate the Email-Sending Tool Use the actual key you configured in "Runtime Variables" as the variable name. For example, you can use email, or a naming style that suits you better such as my company.email. Help me build a small email-sending tool. It should support entering a subject, editing the body, and uploading attachments. Note: get the email account, password, and related information from the email variable. Do not ask the user to enter them manually.

Use Variables to Generate the Email-Sending Tool

examples/excel_rows_stat Excel Row Count Text Example Help me generate an Excel row-count tool. The input is an Excel file path, and the output is JSON containing the total number of rows in the whole file plus the row count for each worksheet. Ignore the header row and any data rows where all columns are empty. Note The tool must be able to handle very large Excel files without crashing due to insufficient memory. Excel Row Count Text Example

examples/excel_rows_stat

Excel Row Count

examples/excel_rows_stat Text Example Help me generate an Excel row-count tool. The input is an Excel file path, and the output is JSON containing the total number of rows in the whole file plus the row count for each worksheet. Ignore the header row and any data rows where all columns are empty. Note The tool must be able to handle very large Excel files without crashing due to insufficient memory.

Text Example

examples/pdf_parse PDF Data Extraction This example shows how to generate a tool, through conversation, that extracts target data from a text-based PDF. Prompt Help me generate a PDF data extraction tool. The input is a PDF file path, and the results are appended to a CSV file. The PDF text content looks roughly like this: Some Company Ltd. Applicant: Zhang San, phone: 13800000000, email: zhangsan@example.com, address: Haidian District, Beijing. Signed a contract with Some Company Ltd. in a certain year and month; the contract number is 2023001. Agreed that Extract the following fields: company name, applicant, phone, email, address, contract number, contract content. PDF Data Extraction Prompt

examples/pdf_parse

examples/pdf_parse PDF Data Extraction This example shows how to generate a tool, through conversation, that extracts target data from a text-based PDF.

PDF Data Extraction

examples/pdf_parse Prompt Help me generate a PDF data extraction tool. The input is a PDF file path, and the results are appended to a CSV file. The PDF text content looks roughly like this: Some Company Ltd. Applicant: Zhang San, phone: 13800000000, email: zhangsan@example.com, address: Haidian District, Beijing. Signed a contract with Some Company Ltd. in a certain year and month; the contract number is 2023001. Agreed that Extract the following fields: company name, applicant, phone, email, address, contract number, contract content.

Prompt

examples/pdf_parse_and_send_mail Extract PDF Data and Send Email After you have generated the relevant tools, you can try chaining several of them together through an AI conversation to accomplish more complex tasks. Text Example Manually drag a file into the chat box. Enter the following AI conversation: Use the PDF data extraction tool to extract data from the PDF file, and use the email-sending tool to send the generated result file to the relevant people. Email subject: PDF data extraction result Email body: Please see the PDF data extraction result file in the attachment Recipients: zhangsan@example.com, lisi@example.com Notes First, make sure you have already generated the relevant tools. If the PDF file contains sensitive information, be sure not to let the tool output results directly as JSON; instead, require it to write the results to a file and send that file as an email attachment. This avoids exposing the PDF content to the AI model during the conversation. Make sure you have configured the account, password, and other sensitive information required by the email-sending tool, and set them up in runtime variables. For variable configuration, see Runtime Variables. Extract PDF Data and Send Email Text Example Notes

examples/pdf_parse_and_send_mail

examples/pdf_parse_and_send_mail Extract PDF Data and Send Email After you have generated the relevant tools, you can try chaining several of them together through an AI conversation to accomplish more complex tasks.

Extract PDF Data and Send Email

examples/pdf_parse_and_send_mail Text Example Manually drag a file into the chat box. Enter the following AI conversation: Use the PDF data extraction tool to extract data from the PDF file, and use the email-sending tool to send the generated result file to the relevant people. Email subject: PDF data extraction result Email body: Please see the PDF data extraction result file in the attachment Recipients: zhangsan@example.com, lisi@example.com

examples/pdf_parse_and_send_mail Notes First, make sure you have already generated the relevant tools. If the PDF file contains sensitive information, be sure not to let the tool output results directly as JSON; instead, require it to write the results to a file and send that file as an email attachment. This avoids exposing the PDF content to the AI model during the conversation. Make sure you have configured the account, password, and other sensitive information required by the email-sending tool, and set them up in runtime variables. For variable configuration, see Runtime Variables.

Notes

How to Describe a Tool How to Have the AI Build a Tool When writing a tool requirement, the point is not to "write a lot"—it is to describe the interface clearly. A good tool prompt usually only needs to answer 6 questions: What does this tool do What is the input What rules to process by What artifact to output What to return in the end Whether to print stage logs Note: documents like this can live in the examples directory; the page path can be index or any other path. Recommended Format Help me build a {tool name} tool. Goal: What task should this tool accomplish Input: What is the input, what type, is it required Processing rules: What to extract, transform, filter, or count Output: Whether to generate a file If a file is generated, specify its format and save location Return: Return JSON Or return the result file path Logging: Whether to print stage logs Constraints: Requirements for large-file handling, error handling, sensitive information, etc. The Difference Between "Output" and "Return" Output: the actual result the tool produces, such as a CSV file, an Excel file, or a piece of structured data Return: what the tool hands back directly when execution finishes, such as the result file path, or JSON Two common patterns: Output a file, and return the path to that file Do not generate a file, and return JSON directly Three Most Common Examples Output a File, Return the File Path Help me build a PDF information extraction tool. The input is a PDF file path. Extract the company name, contact person, phone number, and contract number. Output a CSV file saved in the same directory as the input file. For the return value, provide the full path to the generated CSV file. Please print stage logs for reading the PDF, extracting fields, and writing results. Do Not Generate a File, Return JSON Directly Help me build a web content statistics tool. The input is a web page URL. Extract the plain text of the page body and count characters, paragraphs, and links. Do not generate a file; return JSON directly. Return Both a File Path and Summary Information Help me build a CSV cleaning tool. The input is a CSV file path. Output a new, cleaned CSV file. The return value should include the result file path, the original row count, the cleaned row count, and the number of removed rows. Please print stage logs, but do not output raw sensitive content. A One-Line Reminder When results are large or contain sensitive information, prefer "output a file + return the file path" When results are small and have a fixed structure, prefer returning JSON directly Do not write accounts, passwords, or keys directly into the prompt—read them from runtime variables Self-Check After Writing If you can clearly answer these 5 questions, it is usually enough: What is the input How to process it What is the output What to return Whether logs are needed Further Reading Excel Row-Count Tool PDF Data Extraction Tool Web Content Extraction Tool Email Tool and Variables How to Have the AI Build a Tool Recommended Format The Difference Between "Output" and "Return" Three Most Common Examples 1. Output a File, Return the File Path 2. Do Not Generate a File, Return JSON Directly 3. Return Both a File Path and Summary Information A One-Line Reminder Self-Check After Writing Further Reading

How to Describe a Tool

How to Describe a Tool How to Have the AI Build a Tool When writing a tool requirement, the point is not to "write a lot"—it is to describe the interface clearly. A good tool prompt usually only needs to answer 6 questions: What does this tool do What is the input What rules to process by What artifact to output What to return in the end Whether to print stage logs Note: documents like this can live in the examples directory; the page path can be index or any other path.

How to Have the AI Build a Tool

How to Describe a Tool Recommended Format Help me build a {tool name} tool. Goal: What task should this tool accomplish Input: What is the input, what type, is it required Processing rules: What to extract, transform, filter, or count Output: Whether to generate a file If a file is generated, specify its format and save location Return: Return JSON Or return the result file path Logging: Whether to print stage logs Constraints: Requirements for large-file handling, error handling, sensitive information, etc.

Recommended Format

How to Describe a Tool The Difference Between "Output" and "Return" Output: the actual result the tool produces, such as a CSV file, an Excel file, or a piece of structured data Return: what the tool hands back directly when execution finishes, such as the result file path, or JSON Two common patterns: Output a file, and return the path to that file Do not generate a file, and return JSON directly

The Difference Between "Output" and "Return"

How to Describe a Tool Three Most Common Examples

Three Most Common Examples

How to Describe a Tool 1. Output a File, Return the File Path Help me build a PDF information extraction tool. The input is a PDF file path. Extract the company name, contact person, phone number, and contract number. Output a CSV file saved in the same directory as the input file. For the return value, provide the full path to the generated CSV file. Please print stage logs for reading the PDF, extracting fields, and writing results.

1. Output a File, Return the File Path

How to Describe a Tool 2. Do Not Generate a File, Return JSON Directly Help me build a web content statistics tool. The input is a web page URL. Extract the plain text of the page body and count characters, paragraphs, and links. Do not generate a file; return JSON directly.

2. Do Not Generate a File, Return JSON Directly

How to Describe a Tool 3. Return Both a File Path and Summary Information Help me build a CSV cleaning tool. The input is a CSV file path. Output a new, cleaned CSV file. The return value should include the result file path, the original row count, the cleaned row count, and the number of removed rows. Please print stage logs, but do not output raw sensitive content.

3. Return Both a File Path and Summary Information

How to Describe a Tool A One-Line Reminder When results are large or contain sensitive information, prefer "output a file + return the file path" When results are small and have a fixed structure, prefer returning JSON directly Do not write accounts, passwords, or keys directly into the prompt—read them from runtime variables

A One-Line Reminder

How to Describe a Tool Self-Check After Writing If you can clearly answer these 5 questions, it is usually enough: What is the input How to process it What is the output What to return Whether logs are needed

Self-Check After Writing

How to Describe a Tool Further Reading Excel Row-Count Tool PDF Data Extraction Tool Web Content Extraction Tool Email Tool and Variables

examples/web_content_parse

examples/web_content_parse Web Content Extraction

Web Content Extraction

examples/web_content_parse Text Example This example shows how to generate a tool, through conversation, that extracts target data from a web page. Simple: extract web page content and output it as text Help me generate a web page extraction tool. The tool takes a web page URL as input and outputs the plain text content of the page, with HTML elements removed. Related notes

examples/web_content_parse Notes If you need to extract text content from a web page, you can follow the approach of the PDF Data Extraction tool.

Home What is Jelly Studio Jelly Studio is a desktop AI toolkit built for business workflows. It helps you turn repetitive work—file processing, spreadsheet cleanup, information extraction, data conversion—into reusable tools, and compose them into controllable flows when needed. Home In one sentence: Jelly Studio makes repetitive work simpler. What it can do Turn repetitive steps across spreadsheets, files, receipts, and exported data into tools Run tools locally under control, with explicit directory, network, and runtime variable permissions Let you start a task first, then review and verify the results Reuse existing tools and compose them into longer flows when needed What it does not do It is not an AI IDE; it does not center on coding or engineering collaboration It is not a general-purpose automation platform; it does not try to cover every scenario It is not an open extension marketplace; it does not focus on installing large amounts of external capabilities It is not a fully autonomous agent; by default a human still initiates and reviews Where it fits Spreadsheet cleaning, formatting, and batch statistics Information extraction from PDFs, receipts, and exported files Data conversion, reconciliation, aggregation, and cross-checking Batch generation of standardized emails, reports, or result files Everyday business processing that needs "visible process, checkable results, verifiable execution" Where it does not fit well Open-ended creation, highly subjective judgment, or tasks whose results are hard to verify Fully automated flows that require long autonomous planning, multi-system orchestration, and almost no human review Deep coding workflows aimed at developers Scenarios where an extension ecosystem and plugin marketplace are the top priority How it works Describe a clear, repetitive business action in natural language Let Jelly Studio generate or adjust the corresponding tool Configure the model, runtime variables, and execution permissions for the tool Run it manually and check the results Reuse stable steps repeatedly, and compose them into longer flows when needed Why it suits business processing Controllable : directory, network, and subprocess permissions are configurable, reducing "black-box execution" Reviewable : you can see the run inputs, execution results, and error messages Reusable : a tool built once can run repeatedly, instead of being re-described each time Safer : sensitive information can be managed through runtime variables, without writing it directly into the conversation Quick start Download and install Jelly Studio Configure a model Complete licensing Create and run your first tool Learn how to describe a tool Download Click the link below to open the download directory and choose the build that fits your system. Platform Requirement File Name Windows Windows 10/11 Jelly Studio-x.y.z-windows.exe macOS (Apple Silicon) macOS 10.15+ Jelly Studio-darwin-arm64-x.y.z.zip macOS (Intel) macOS 10.15+ Jelly Studio-darwin-x64-x.y.z.zip Download link: Baidu Netdisk Access code: ygcv What is Jelly Studio What it can do What it does not do Where it fits Where it does not fit well How it works Why it suits business processing Quick start Download

Home

Home What is Jelly Studio Jelly Studio is a desktop AI toolkit built for business workflows. It helps you turn repetitive work—file processing, spreadsheet cleanup, information extraction, data conversion—into reusable tools, and compose them into controllable flows when needed. Home In one sentence: Jelly Studio makes repetitive work simpler.

What is Jelly Studio

Home What it can do Turn repetitive steps across spreadsheets, files, receipts, and exported data into tools Run tools locally under control, with explicit directory, network, and runtime variable permissions Let you start a task first, then review and verify the results Reuse existing tools and compose them into longer flows when needed

What it can do

Home What it does not do It is not an AI IDE; it does not center on coding or engineering collaboration It is not a general-purpose automation platform; it does not try to cover every scenario It is not an open extension marketplace; it does not focus on installing large amounts of external capabilities It is not a fully autonomous agent; by default a human still initiates and reviews

What it does not do

Home Where it fits Spreadsheet cleaning, formatting, and batch statistics Information extraction from PDFs, receipts, and exported files Data conversion, reconciliation, aggregation, and cross-checking Batch generation of standardized emails, reports, or result files Everyday business processing that needs "visible process, checkable results, verifiable execution"

Where it fits

Home Where it does not fit well Open-ended creation, highly subjective judgment, or tasks whose results are hard to verify Fully automated flows that require long autonomous planning, multi-system orchestration, and almost no human review Deep coding workflows aimed at developers Scenarios where an extension ecosystem and plugin marketplace are the top priority

Where it does not fit well

Home How it works Describe a clear, repetitive business action in natural language Let Jelly Studio generate or adjust the corresponding tool Configure the model, runtime variables, and execution permissions for the tool Run it manually and check the results Reuse stable steps repeatedly, and compose them into longer flows when needed

How it works

Home Why it suits business processing Controllable : directory, network, and subprocess permissions are configurable, reducing "black-box execution" Reviewable : you can see the run inputs, execution results, and error messages Reusable : a tool built once can run repeatedly, instead of being re-described each time Safer : sensitive information can be managed through runtime variables, without writing it directly into the conversation

Why it suits business processing

Home Quick start Download and install Jelly Studio Configure a model Complete licensing Create and run your first tool Learn how to describe a tool

Quick start

Home Download Click the link below to open the download directory and choose the build that fits your system. Platform Requirement File Name Windows Windows 10/11 Jelly Studio-x.y.z-windows.exe macOS (Apple Silicon) macOS 10.15+ Jelly Studio-darwin-arm64-x.y.z.zip macOS (Intel) macOS 10.15+ Jelly Studio-darwin-x64-x.y.z.zip Download link: Baidu Netdisk Access code: ygcv

Download

Install Installation Guide This guide walks you through installing Jelly Studio. System Requirements Before you begin, make sure your system meets the following requirements: Windows : Windows 10 or later macOS : macOS 10.15 (Catalina) or later Download and Install Click the link below to open the download directory and choose the build that fits your system. Platform Requirement File Name Windows Windows 10/11 Jelly Studio-x.y.z-windows.exe macOS (Apple Silicon) macOS 10.15+ Jelly Studio-darwin-arm64-x.y.z.zip macOS (Intel) macOS 10.15+ Jelly Studio-darwin-x64-x.y.z.zip Download link: Baidu Netdisk Access code: ygcv Windows Installation Download the .exe installer. Double-click to run the installer. Follow the setup wizard to complete the installation. macOS Installation Download the matching .zip file. Unzip the file. Drag the app into the Applications folder. Tip : On first launch, macOS may warn that "the developer cannot be verified". Go to System Settings > Privacy & Security and allow it to run. Installation Guide System Requirements Download and Install Windows Installation macOS Installation

Install

Install Installation Guide This guide walks you through installing Jelly Studio.

Installation Guide

Install System Requirements Before you begin, make sure your system meets the following requirements: Windows : Windows 10 or later macOS : macOS 10.15 (Catalina) or later

System Requirements

Install Download and Install Click the link below to open the download directory and choose the build that fits your system. Platform Requirement File Name Windows Windows 10/11 Jelly Studio-x.y.z-windows.exe macOS (Apple Silicon) macOS 10.15+ Jelly Studio-darwin-arm64-x.y.z.zip macOS (Intel) macOS 10.15+ Jelly Studio-darwin-x64-x.y.z.zip Download link: Baidu Netdisk Access code: ygcv

Download and Install

Install Windows Installation Download the .exe installer. Double-click to run the installer. Follow the setup wizard to complete the installation.

Windows Installation

Install macOS Installation Download the matching .zip file. Unzip the file. Drag the app into the Applications folder. Tip : On first launch, macOS may warn that "the developer cannot be verified". Go to System Settings > Privacy & Security and allow it to run.

macOS Installation

Model Configuration Model Configuration This guide walks you through configuring an AI model so Jelly Studio can work properly. Model Recommendation We recommend starting with DeepSeek models—they offer strong value for everyday use. Tip : For your first try, top up a small amount (e.g. a few dollars) to experiment and avoid unnecessary spending. DeepSeek official site Supported Models Jelly Studio supports the following types of models: DeepSeek : the full model lineup (primary target for adaptation and testing). Alibaba Cloud Tongyi Qianwen : the Tongyi Qianwen model series. Other models : any model compatible with the OpenAI API can be used. Configuration Steps Open the app, go to Settings > Model , and configure the parameters as shown below. Model configuration Getting an API Key For how to obtain an API Key, refer to the official documentation of the LLM platform you use, or contact its support team for help. Model Configuration Model Recommendation Supported Models Configuration Steps Getting an API Key

Model Configuration

Model Configuration Model Recommendation We recommend starting with DeepSeek models—they offer strong value for everyday use. Tip : For your first try, top up a small amount (e.g. a few dollars) to experiment and avoid unnecessary spending. DeepSeek official site

Model Recommendation

Model Configuration Supported Models Jelly Studio supports the following types of models: DeepSeek : the full model lineup (primary target for adaptation and testing). Alibaba Cloud Tongyi Qianwen : the Tongyi Qianwen model series. Other models : any model compatible with the OpenAI API can be used.

Supported Models

Model Configuration Configuration Steps Open the app, go to Settings > Model , and configure the parameters as shown below. Model configuration

Configuration Steps

Model Configuration Getting an API Key For how to obtain an API Key, refer to the official documentation of the LLM platform you use, or contact its support team for help.

Getting an API Key

Run Running a Tool Start a conversation and let the AI help you generate a practical tool. Create Your First Tool Create a tool Steps Describe your need : type your requirement in the chat box on the right. Help me generate an Excel row-count tool. The input is an Excel file, and the output is JSON containing the total number of rows in the file and the number of rows in each sheet. Skip the header row and any data rows where all columns are empty. Wait for generation : the AI analyzes your requirement and generates the corresponding tool code. Run the tool : Find the tool list in the feature center and double-click the tool you just created. Enter parameters on the tool run page and click Run to view the results. Tool list Tool run Error Handling If a tool fails during execution, you can handle it as follows. Error handling Steps Copy the error message. In the chat box, type: tool execution error + the error message. The AI analyzes the error and suggests a fix. Tip : If you started a new conversation, you can associate a tool in these ways: Manually specify the tool name to reference the failing tool. Drag the tool into the chat box to link it to the current conversation. Running a Tool Create Your First Tool Steps Error Handling Steps

Run Running a Tool Start a conversation and let the AI help you generate a practical tool.

Running a Tool

Create Your First Tool

Run Steps Describe your need : type your requirement in the chat box on the right. Help me generate an Excel row-count tool. The input is an Excel file, and the output is JSON containing the total number of rows in the file and the number of rows in each sheet. Skip the header row and any data rows where all columns are empty. Wait for generation : the AI analyzes your requirement and generates the corresponding tool code. Run the tool : Find the tool list in the feature center and double-click the tool you just created. Enter parameters on the tool run page and click Run to view the results. Tool list Tool run

Steps

Run Error Handling If a tool fails during execution, you can handle it as follows. Error handling

Error Handling

Run Steps Copy the error message. In the chat box, type: tool execution error + the error message. The AI analyzes the error and suggests a fix. Tip : If you started a new conversation, you can associate a tool in these ways: Manually specify the tool name to reference the failing tool. Drag the tool into the chat box to link it to the current conversation.

support/contact Contact & Support Contact WeChat Official Account: 玄都洞主 Feedback We have not yet invested in building an automated feedback system. If you run into any issues, please reach out through our WeChat Official Account. Contact & Support Contact Feedback

support/contact

Contact & Support

support/contact Contact WeChat Official Account: 玄都洞主

Contact

support/contact Feedback We have not yet invested in building an automated feedback system. If you run into any issues, please reach out through our WeChat Official Account.

Feedback


# 运行时变量

运行时变量用于管理隐私信息，避免将敏感信息直接写入技能代码或暴露给 LLM。

## 什么是运行时变量

当技能需要使用敏感信息（如 API Key、账号密码等）时，请优先使用运行时变量功能。**切勿在与 AI 的对话中直接输入敏感信息。**

您也可以将运行时变量作为通用配置功能使用，方便在不同技能间共享配置。

## 使用指南

### 1. 配置变量

打开应用，进入 **设置 > 运行时变量**。

![运行时变量](/images/jelly-studio/runtime_vars.png)

在此页面您可以：
- 添加新变量。
- 编辑已有变量。
- 删除不再使用的变量。

### 2. 在对话中使用

在与 LLM 对话时，直接告诉 AI 使用某个变量即可，工具会自动搜索并检查该变量是否存在。

```txt
帮我开发一个邮件发送工具，该工具支持标题填写、邮件正文编辑、邮件附件上传。

说明：邮件相关的账号密码等参数从 email 变量中获取，不要要求用户输入。
```

![用户变量在对话中](/images/jelly-studio/user_vars_in_chat.png)

### 3. 引用变量

在对话中引用变量时，AI 会自动识别并读取对应的配置值，无需手动输入敏感信息。

> **安全提示**：运行时变量存储在本地，不会上传到云端，也不会发送给 LLM，从而更好地保障您的信息安全。


运行时变量 运行时变量 运行时变量用于管理隐私信息，避免将敏感信息直接写入技能代码或暴露给 LLM。 什么是运行时变量 当技能需要使用敏感信息（如 API Key、账号密码等）时，请优先使用运行时变量功能。 切勿在与 AI 的对话中直接输入敏感信息。 您也可以将运行时变量作为通用配置功能使用，方便在不同技能间共享配置。 使用指南 配置变量 打开应用，进入 设置 > 运行时变量 。 运行时变量 在此页面您可以： 添加新变量。 编辑已有变量。 删除不再使用的变量。 在对话中使用 在与 LLM 对话时，直接告诉 AI 使用某个变量即可，工具会自动搜索并检查该变量是否存在。 帮我开发一个邮件发送工具，该工具支持标题填写、邮件正文编辑、邮件附件上传。 说明：邮件相关的账号密码等参数从 email 变量中获取，不要要求用户输入。 用户变量在对话中 引用变量 在对话中引用变量时，AI 会自动识别并读取对应的配置值，无需手动输入敏感信息。 安全提示 ：运行时变量存储在本地，不会上传到云端，也不会发送给 LLM，从而更好地保障您的信息安全。 运行时变量 什么是运行时变量 使用指南 1. 配置变量 2. 在对话中使用 3. 引用变量

运行时变量

运行时变量 什么是运行时变量 当技能需要使用敏感信息（如 API Key、账号密码等）时，请优先使用运行时变量功能。 切勿在与 AI 的对话中直接输入敏感信息。 您也可以将运行时变量作为通用配置功能使用，方便在不同技能间共享配置。

什么是运行时变量

使用指南

运行时变量 1. 配置变量 打开应用，进入 设置 > 运行时变量 。 运行时变量 在此页面您可以： 添加新变量。 编辑已有变量。 删除不再使用的变量。

1. 配置变量

运行时变量 2. 在对话中使用 在与 LLM 对话时，直接告诉 AI 使用某个变量即可，工具会自动搜索并检查该变量是否存在。 帮我开发一个邮件发送工具，该工具支持标题填写、邮件正文编辑、邮件附件上传。 说明：邮件相关的账号密码等参数从 email 变量中获取，不要要求用户输入。 用户变量在对话中

2. 在对话中使用

运行时变量 3. 引用变量 在对话中引用变量时，AI 会自动识别并读取对应的配置值，无需手动输入敏感信息。 安全提示 ：运行时变量存储在本地，不会上传到云端，也不会发送给 LLM，从而更好地保障您的信息安全。

3. 引用变量

examples/email_send 邮件发送工具与变量使用 邮件发送工具通常会涉及敏感信息。请不要在与 AI 的对话中直接提供邮箱账号、密码等内容。建议您先在运行时变量中创建相关变量，并在变量中配置所需参数。关于变量配置，请参考 运行时变量。 文本示例 通过 AI 对话梳理邮件发送所需变量字段 我想开发一个邮件发送工具。通常一封邮件需要配置哪些信息？哪些内容适合放在变量里，哪些内容适合在运行时作为参数输入？如果作为变量，一般字段应该如何命名，尤其是英文命名该怎么设计？ 使用变量生成邮件发送工具 变量名请使用您在“运行时变量”中实际配置的 key。例如可以使用 email，也可以使用 my company.email 这类更符合自己习惯的命名。 帮我开发一个邮件发送小工具。该工具需要支持邮件标题填写、正文编辑和附件上传。 说明：邮件相关的账号、密码等信息从 email 变量中获取，不要要求用户手动输入。 邮件发送工具与变量使用 文本示例 通过 AI 对话梳理邮件发送所需变量字段 使用变量生成邮件发送工具

examples/email_send 邮件发送工具与变量使用 邮件发送工具通常会涉及敏感信息。请不要在与 AI 的对话中直接提供邮箱账号、密码等内容。建议您先在运行时变量中创建相关变量，并在变量中配置所需参数。关于变量配置，请参考 运行时变量。

邮件发送工具与变量使用

文本示例

examples/email_send 通过 AI 对话梳理邮件发送所需变量字段 我想开发一个邮件发送工具。通常一封邮件需要配置哪些信息？哪些内容适合放在变量里，哪些内容适合在运行时作为参数输入？如果作为变量，一般字段应该如何命名，尤其是英文命名该怎么设计？

通过 AI 对话梳理邮件发送所需变量字段

examples/email_send 使用变量生成邮件发送工具 变量名请使用您在“运行时变量”中实际配置的 key。例如可以使用 email，也可以使用 my company.email 这类更符合自己习惯的命名。 帮我开发一个邮件发送小工具。该工具需要支持邮件标题填写、正文编辑和附件上传。 说明：邮件相关的账号、密码等信息从 email 变量中获取，不要要求用户手动输入。

使用变量生成邮件发送工具

examples/excel_rows_stat Excel 行数统计 文本示例 帮我生成一个 Excel 行数统计工具。输入为一个 Excel 文件路径，输出为 JSON 格式，包含整个文件的总行数以及每个工作表的行数统计。请忽略表头行，以及所有列都为空的数据行。 说明 该工具需要支持处理非常大的 Excel 文件，不能因为内存不足而崩溃。 Excel 行数统计 文本示例

examples/excel_rows_stat Excel 行数统计

Excel 行数统计

examples/excel_rows_stat 文本示例 帮我生成一个 Excel 行数统计工具。输入为一个 Excel 文件路径，输出为 JSON 格式，包含整个文件的总行数以及每个工作表的行数统计。请忽略表头行，以及所有列都为空的数据行。 说明 该工具需要支持处理非常大的 Excel 文件，不能因为内存不足而崩溃。

examples/pdf_parse PDF 数据提取 本示例演示如何通过对话生成一个从文本型 PDF 中提取目标数据的工具。 提示词 帮我生成一个 PDF 数据提取工具。输入为一个 PDF 文件路径，输出结果以追加方式写入一个 CSV 文件中。PDF 文本内容格式大致如下： 某某某有限公司 申请人: 张三，手机号: 13800000000，邮箱: zhangsan@example.com，联系地址: 北京市海淀区。于某年某月与某某某有限公司签订合同，合同编号为: 2023001。约定 要求提取以下字段：公司名称、申请人、手机号、邮箱、联系地址、合同编号、合同内容。 PDF 数据提取 提示词

examples/pdf_parse PDF 数据提取 本示例演示如何通过对话生成一个从文本型 PDF 中提取目标数据的工具。

PDF 数据提取

examples/pdf_parse 提示词 帮我生成一个 PDF 数据提取工具。输入为一个 PDF 文件路径，输出结果以追加方式写入一个 CSV 文件中。PDF 文本内容格式大致如下： 某某某有限公司 申请人: 张三，手机号: 13800000000，邮箱: zhangsan@example.com，联系地址: 北京市海淀区。于某年某月与某某某有限公司签订合同，合同编号为: 2023001。约定 要求提取以下字段：公司名称、申请人、手机号、邮箱、联系地址、合同编号、合同内容。

提示词

examples/pdf_parse_and_send_mail PDF 数据提取并发送邮件 当您生成相关工具后，可以尝试通过 AI 对话将多个工具串联起来，完成更复杂的任务。 ## 文本示例 手动将文件拖入对话框中。 输入以下 AI 对话内容： 使用 PDF 数据提取工具提取 PDF 文件中的数据，并通过邮件发送工具将生成的结果文件发送给相关人员。 邮件标题为: PDF 数据提取结果 邮件正文为: 请查看附件中的 PDF 数据提取结果文件 邮件接收人: zhangsan@example.com, lisi@example.com ## 说明 首先请确保您已经生成了相关工具。若 PDF 文件中包含敏感信息，务必不要让工具以 JSON 形式直接输出结果，而应要求其将结果写入文件中，再通过邮件工具以附件形式发送。这样可以避免 PDF 内容在对话过程中暴露给 AI 模型。 请确保您已经配置好邮件发送工具所需的账号、密码等敏感信息，并已在运行时变量中完成相关设置。关于变量配置，请参考 运行时变量。 PDF 数据提取并发送邮件

examples/pdf_parse_and_send_mail PDF 数据提取并发送邮件 当您生成相关工具后，可以尝试通过 AI 对话将多个工具串联起来，完成更复杂的任务。 ## 文本示例 手动将文件拖入对话框中。 输入以下 AI 对话内容： 使用 PDF 数据提取工具提取 PDF 文件中的数据，并通过邮件发送工具将生成的结果文件发送给相关人员。 邮件标题为: PDF 数据提取结果 邮件正文为: 请查看附件中的 PDF 数据提取结果文件 邮件接收人: zhangsan@example.com, lisi@example.com ## 说明 首先请确保您已经生成了相关工具。若 PDF 文件中包含敏感信息，务必不要让工具以 JSON 形式直接输出结果，而应要求其将结果写入文件中，再通过邮件工具以附件形式发送。这样可以避免 PDF 内容在对话过程中暴露给 AI 模型。 请确保您已经配置好邮件发送工具所需的账号、密码等敏感信息，并已在运行时变量中完成相关设置。关于变量配置，请参考 运行时变量。

PDF 数据提取并发送邮件

如何描述一个工具 如何让 AI 开发一个工具 写工具需求时，重点不是“描述得很长”，而是把接口说清楚。 一个好用的工具提示词，通常只需要回答 6 个问题： 这个工具做什么 输入是什么 按什么规则处理 输出什么产物 最后返回什么 是否需要打印阶段日志 说明：这类说明文档放在 examples 目录即可，页面路径用 index 或其他路径都可以。 推荐写法 帮我开发一个{工具名称}工具。 工具目标： 这个工具要完成什么任务 输入： 输入是什么，类型是什么，是否必填 处理规则： 需要提取、转换、过滤、统计什么内容 输出： 是否生成文件 如果生成文件，说明文件格式和保存位置 返回： 返回 JSON 或返回结果文件路径 日志： 是否打印阶段日志 约束： 大文件处理、异常处理、敏感信息处理等要求 “输出”和“返回”的区别 输出：工具真正生成的结果，比如一个 CSV 文件、Excel 文件，或者一段结构化数据 返回：工具执行结束后直接交还的内容，比如结果文件路径，或者 JSON 常见的两种写法： 输出一个文件，返回这个文件的路径 不生成文件，直接返回 JSON 三个最常见的例子 输出文件，返回文件路径 帮我开发一个 PDF 信息提取工具。 输入为一个 PDF 文件路径。 提取公司名称、联系人、手机号和合同编号。 输出为一个 CSV 文件，保存在输入文件同级目录。 返回值请提供生成后的 CSV 文件完整路径。 请打印读取 PDF、提取字段、写入结果这几个阶段日志。 不生成文件，直接返回 JSON 帮我开发一个网页内容统计工具。 输入为一个网页地址。 提取网页正文纯文本，并统计字符数、段落数和链接数。 不要生成文件，直接返回 JSON。 既返回文件路径，也返回摘要信息 帮我开发一个 CSV 清洗工具。 输入为一个 CSV 文件路径。 输出为一个清洗后的新 CSV 文件。 返回值请包含结果文件路径、原始行数、清洗后行数和删除行数。 请打印阶段日志，但不要输出原始敏感内容。 一句话提醒 结果很多或包含敏感信息时，优先“输出文件 + 返回文件路径” 结果很小、结构固定时，优先直接返回 JSON 账号、密码、密钥不要直接写进提示词，应从运行时变量读取 写完后自查一下 如果这 5 个问题都能回答清楚，通常就够了： 输入是什么 怎么处理 输出是什么 返回什么 要不要日志 延伸阅读 Excel 行数统计工具 PDF 数据提取工具 网页内容提取工具 邮件发送工具及变量使用 如何让 AI 开发一个工具 推荐写法 “输出”和“返回”的区别 三个最常见的例子 1. 输出文件，返回文件路径 2. 不生成文件，直接返回 JSON 3. 既返回文件路径，也返回摘要信息 一句话提醒 写完后自查一下 延伸阅读

如何描述一个工具

如何描述一个工具 如何让 AI 开发一个工具 写工具需求时，重点不是“描述得很长”，而是把接口说清楚。 一个好用的工具提示词，通常只需要回答 6 个问题： 这个工具做什么 输入是什么 按什么规则处理 输出什么产物 最后返回什么 是否需要打印阶段日志 说明：这类说明文档放在 examples 目录即可，页面路径用 index 或其他路径都可以。

如何让 AI 开发一个工具

如何描述一个工具 推荐写法 帮我开发一个{工具名称}工具。 工具目标： 这个工具要完成什么任务 输入： 输入是什么，类型是什么，是否必填 处理规则： 需要提取、转换、过滤、统计什么内容 输出： 是否生成文件 如果生成文件，说明文件格式和保存位置 返回： 返回 JSON 或返回结果文件路径 日志： 是否打印阶段日志 约束： 大文件处理、异常处理、敏感信息处理等要求

“输出”和“返回”的区别

如何描述一个工具 三个最常见的例子

三个最常见的例子

如何描述一个工具 1. 输出文件，返回文件路径 帮我开发一个 PDF 信息提取工具。 输入为一个 PDF 文件路径。 提取公司名称、联系人、手机号和合同编号。 输出为一个 CSV 文件，保存在输入文件同级目录。 返回值请提供生成后的 CSV 文件完整路径。 请打印读取 PDF、提取字段、写入结果这几个阶段日志。

1. 输出文件，返回文件路径

如何描述一个工具 2. 不生成文件，直接返回 JSON 帮我开发一个网页内容统计工具。 输入为一个网页地址。 提取网页正文纯文本，并统计字符数、段落数和链接数。 不要生成文件，直接返回 JSON。

2. 不生成文件，直接返回 JSON

如何描述一个工具 3. 既返回文件路径，也返回摘要信息 帮我开发一个 CSV 清洗工具。 输入为一个 CSV 文件路径。 输出为一个清洗后的新 CSV 文件。 返回值请包含结果文件路径、原始行数、清洗后行数和删除行数。 请打印阶段日志，但不要输出原始敏感内容。

3. 既返回文件路径，也返回摘要信息

如何描述一个工具 一句话提醒 结果很多或包含敏感信息时，优先“输出文件 + 返回文件路径” 结果很小、结构固定时，优先直接返回 JSON 账号、密码、密钥不要直接写进提示词，应从运行时变量读取

一句话提醒

如何描述一个工具 写完后自查一下 如果这 5 个问题都能回答清楚，通常就够了： 输入是什么 怎么处理 输出是什么 返回什么 要不要日志

写完后自查一下

如何描述一个工具 延伸阅读 Excel 行数统计工具 PDF 数据提取工具 网页内容提取工具 邮件发送工具及变量使用

延伸阅读

examples/web_content_parse 网页内容提取 文本示例 本示例演示如何通过对话生成一个从网页中提取目标数据的工具。 简单提取网页内容并输出为文本 帮我生成一个 Web 网页提取工具。该工具接收一个网页地址作为输入，并输出网页的纯文本内容，移除 HTML 元素。 相关说明 说明 如果您需要从网页中提取文本内容，可以参考 PDF 数据提取 工具的思路。 网页内容提取 文本示例 说明

examples/web_content_parse 网页内容提取

网页内容提取

examples/web_content_parse 文本示例 本示例演示如何通过对话生成一个从网页中提取目标数据的工具。 简单提取网页内容并输出为文本 帮我生成一个 Web 网页提取工具。该工具接收一个网页地址作为输入，并输出网页的纯文本内容，移除 HTML 元素。 相关说明

examples/web_content_parse 说明 如果您需要从网页中提取文本内容，可以参考 PDF 数据提取 工具的思路。

说明

首页 Jelly Studio 是什么 Jelly Studio 是一个面向业务处理场景的桌面 AI 工具平台。 它帮助您把文件处理、表格整理、信息提取、数据转换这类重复工作做成可复用工具，并在需要时组合成可控流程。 首页 如果只用一句话说明：Jelly Studio 让重复工作更简单。 它能做什么 把表格、文件、票据、导出数据中的重复处理步骤做成工具 让工具在本地受控运行，明确目录、网络和运行变量权限 让用户先发起任务，再查看和复核结果 复用已有工具，并在需要时组合成更长流程 它不做什么 不是 AI IDE，不以写代码和工程协作为核心 不是通用自动化平台，不追求覆盖所有场景 不是开放扩展市场，不以安装海量外部能力为重点 不是全自动代理，默认仍由人发起和复核 适合什么场景 表格清洗、格式整理、批量统计 PDF、票据、导出文件中的信息提取 数据转换、对账、汇总、交叉核对 标准化邮件、报表或结果文件的批量生成 需要“过程可见、结果可查、执行可复核”的日常业务处理 不太适合什么场景 开放式创作、强主观判断、结果难以校验的任务 需要长时间自主规划、多系统联动、几乎无人审核的全自动流程 面向开发者的深度编码工作流 以扩展生态和插件市场为第一优先级的场景 它如何工作 用自然语言描述一个明确、重复的业务动作 让 Jelly Studio 生成或调整对应工具 为工具配置模型、运行时变量和执行权限 手动运行并检查结果 将稳定步骤反复复用，必要时再组合成更长流程 为什么适合业务处理 可控 ：目录、网络、子进程权限可配置，减少“黑盒执行” 可复查 ：用户可以看到运行输入、执行结果和报错信息 可复用 ：一次做好的工具可以反复运行，而不是每次重新描述 更安全 ：敏感信息可以通过运行时变量管理，不必直接写进对话 快速开始 下载安装 Jelly Studio 配置模型 完成授权 创建并运行第一个工具 查看如何描述一个工具 下载 点击下方链接打开下载目录，选择适合您系统的版本进行下载。 平台 版本要求 文件名 Windows Windows 10/11 Jelly Studio-x.y.z-windows.exe macOS (Apple Silicon) macOS 10.15+ Jelly Studio-darwin-arm64-x.y.z.zip macOS (Intel) macOS 10.15+ Jelly Studio-darwin-x64-x.y.z.zip 下载链接： 百度云网盘 提取码： ygcv Jelly Studio 是什么 它能做什么 它不做什么 适合什么场景 不太适合什么场景 它如何工作 为什么适合业务处理 快速开始 下载

首页

首页 Jelly Studio 是什么 Jelly Studio 是一个面向业务处理场景的桌面 AI 工具平台。 它帮助您把文件处理、表格整理、信息提取、数据转换这类重复工作做成可复用工具，并在需要时组合成可控流程。 首页 如果只用一句话说明：Jelly Studio 让重复工作更简单。

Jelly Studio 是什么

首页 它能做什么 把表格、文件、票据、导出数据中的重复处理步骤做成工具 让工具在本地受控运行，明确目录、网络和运行变量权限 让用户先发起任务，再查看和复核结果 复用已有工具，并在需要时组合成更长流程

它能做什么

首页 它不做什么 不是 AI IDE，不以写代码和工程协作为核心 不是通用自动化平台，不追求覆盖所有场景 不是开放扩展市场，不以安装海量外部能力为重点 不是全自动代理，默认仍由人发起和复核

它不做什么

首页 适合什么场景 表格清洗、格式整理、批量统计 PDF、票据、导出文件中的信息提取 数据转换、对账、汇总、交叉核对 标准化邮件、报表或结果文件的批量生成 需要“过程可见、结果可查、执行可复核”的日常业务处理

适合什么场景

首页 不太适合什么场景 开放式创作、强主观判断、结果难以校验的任务 需要长时间自主规划、多系统联动、几乎无人审核的全自动流程 面向开发者的深度编码工作流 以扩展生态和插件市场为第一优先级的场景

不太适合什么场景

首页 它如何工作 用自然语言描述一个明确、重复的业务动作 让 Jelly Studio 生成或调整对应工具 为工具配置模型、运行时变量和执行权限 手动运行并检查结果 将稳定步骤反复复用，必要时再组合成更长流程

它如何工作

首页 为什么适合业务处理 可控 ：目录、网络、子进程权限可配置，减少“黑盒执行” 可复查 ：用户可以看到运行输入、执行结果和报错信息 可复用 ：一次做好的工具可以反复运行，而不是每次重新描述 更安全 ：敏感信息可以通过运行时变量管理，不必直接写进对话