logo
Development
OverviewDevSpaceWorkSpaceAPI ReferenceBest PracticeChangelog
Search
English
WorkSpace / App Settings / Tools
Tool Management
Last updated：4 months ago

Tool Management

Overview

Tools are the "hands and feet" that an Agent uses to carry out actions. The system provides 19 built-in tools and supports extending the toolset via the MCP protocol, OpenAPI definitions, and Skills. The Tool Management page is used to configure and manage all available tools.

Screenshot placeholder

19 Built-in Tools

File Operations (7)

Tool Function Security Level
Read File Reads file content, supports line-number ranges, up to 5MB P0 (auto-executed)
Write File Writes files, automatically creating the file and parent directories if they don't exist P0
Edit Makes local modifications to an existing file (search-and-replace mode) P0
MultiEdit Makes multiple batch modifications to an existing file P0
List Directory Lists files and subdirectories in a directory, showing file sizes P0 (locked)
Glob Files Searches for files by glob pattern, up to 100 results P0 (locked)
Search Files Searches file content by regular expression, up to 250 results P0 (locked)

Command Execution (1)

Tool Function Security Level
Bash Executes shell commands, scripts, and Git operations, 120-second timeout, output up to 50KB P1 (confirm on first use)

Network (1)

Tool Function Security Level
Web Fetch Fetches URL content and converts it to Markdown, with automatic noise removal and 15-minute caching P1

Task Management (1)

Tool Function Security Level
Task Manage Manages a structured task list (to-do / in-progress / done), automatically invoked for tasks with more than 3 steps P0 (locked)

Skill Invocation (1)

Tool Function Security Level
Skill Invokes an installed Skill module to perform a specialized task P1

User Interaction (1)

Tool Function Security Level
AskUserQuestion The Agent asks the user a question (up to 4 questions, each with 2-10 options) P0 (locked)

Sub-agents (3)

Tool Function Security Level
create_local_sub_agent Creates a temporary sub-agent on the current node, supporting 1-3 levels of nesting P0
dispatch_sub_agent Dispatches a task to a registered sub-agent P0
dispatch_multi_node_agent Dispatches a task to a remote node via the Gateway P0

Memory (4)

Tool Function Security Level
Conversation Search Searches historical conversations by keyword, supports time-range filtering, up to 10 results P0 (locked)
Recent Chats Lists the 1-20 most recent conversation sessions, supports time filtering P0 (locked)
Memory Query Searches stored user memory (semantic search + graph expansion) P0 (locked)
Memory Manage Adds, updates, or deletes user memory P1

Security Levels Explained

Level Meaning Behavior
P0 Safe operation Auto-executed, no user confirmation required
P1 Requires confirmation A permission prompt appears on first invocation
P2 Confirm every time User confirmation is required for every invocation
Locked Cannot be downgraded The security level is fixed and cannot be adjusted

Administrators can adjust the permission levels of non-locked tools (the organization's lower bound) in Security Guardrails, and individuals can further tighten them in the APP's runtime security.

MCP Tool Protocol

MCP (Model Context Protocol) is a standardized communication protocol for AI tools that supports extending the toolset through external services.

Three Transport Methods

Transport Description Applicable Scenario
stdio Local subprocess communication (command + args) Local CLI tools
SSE Server-Sent Events (long-lived HTTP connection) Remote services
HTTP Streamable HTTP API Remote services

Five Authentication Modes

Authentication Description
none No authentication
bearer Bearer Token
basic HTTP Basic Auth
service Service-to-service authentication
oauth2 OAuth2 authorization code flow (auto-discovers the well-known endpoint)

MCP Management Interface

Manage MCP servers through the McpManager component:

  • Add server: Configure the name, transport method, connection parameters, and authentication method
  • Edit/Delete: Modify or remove an existing configuration
  • Enable/Disable: Toggle a server's availability
  • Tool discovery: Automatically discover the list of tools provided by a server after connecting

Security Isolation

MCP subprocesses only pass whitelisted environment variables, preventing the leakage of sensitive information such as API Keys.

API Tools

Supports automatically parsing and creating tools from OpenAPI/Swagger documents:

Function Description
Automatic parsing Extracts endpoints and parameter schemas from an OpenAPI specification
Authentication config Five authentication modes
Template variables Dynamic {{variable}} substitution
HTTP methods GET / POST / PUT / DELETE / PATCH

Custom Tool Directory NEW

The system adds a new ToolDirectoryManager — it manages custom tools under the {dataDir}/TOOLs/{orgId}/_custom/ directory and can load them without restarting the APP.

Directory Structure

{dataDir}/TOOLs/{orgId}/_custom/ ├── my-tool-1/ │ ├── metadata.json │ └── ... └── my-tool-2/ ├── metadata.json └── ...
                      
                      {dataDir}/TOOLs/{orgId}/_custom/
├── my-tool-1/
│   ├── metadata.json
│   └── ...
└── my-tool-2/
    ├── metadata.json
    └── ...
This code block in the floating window

metadata.json Example

{ "id": "my-tool-1", "name": "My Custom Tool", "description": "自定义工具说明", "type": "API", "enabled": true, "createdAt": 1776921208476, "updatedAt": 1776921208476 }
                      
                      {
  "id": "my-tool-1",
  "name": "My Custom Tool",
  "description": "自定义工具说明",
  "type": "API",
  "enabled": true,
  "createdAt": 1776921208476,
  "updatedAt": 1776921208476
}
This code block in the floating window

Loading Mechanism

  • Scan method: scanAndLoad(orgId)
  • Type support: API / MCP
  • Organization isolation: separate directories per orgId, with no interference between them
  • Hot loading: no APP restart required

Tool Loading Pipeline

All tools are loaded and merged in the following order:

① Built-in tools (19) ↓ ② MCP tools (discovered from configured MCP servers) ↓ ③ API tools (parsed from OpenAPI definitions) ↓ ④ Custom tools (ToolDirectoryManager scans the _custom/ directory) ↓ ⑤ Skill tools (execute_skill_{id} injected from enabled Skills) ↓ ⑥ Merge and deduplicate → final tool list → sent to the LLM
                      
                      ① Built-in tools (19)
    ↓
② MCP tools (discovered from configured MCP servers)
    ↓
③ API tools (parsed from OpenAPI definitions)
    ↓
④ Custom tools (ToolDirectoryManager scans the _custom/ directory)
    ↓
⑤ Skill tools (execute_skill_{id} injected from enabled Skills)
    ↓
⑥ Merge and deduplicate → final tool list → sent to the LLM
This code block in the floating window

Common Tool Usage Scenarios

Example User Request Tool Called by the Agent Description
"Show me the contents of package.json" Read File Reads and returns the file content
"Rename the function from getData to fetchData" Edit Precise search and replace
"Create a README.md" Write File Creates a new file
"Which .tsx files are in the project?" Glob Files Searches filenames by pattern
"Where is useEffect used in the code?" Search Files Regex search within file content
"Run npm install for me" Bash Executes a shell command
"Summarize the content of this web page for me: https://..." Web Fetch Fetches the web page and converts it to Markdown
"This task is fairly complex, let's outline a plan first" Task Manage Creates a structured task list
"Search for yesterday's conversation about database migration" Conversation Search Searches historical conversations
"Remember that I prefer the dark theme" Memory Manage Stores it to memory

How-to: Add an MCP Tool

  1. Open APP → Settings → Tools → the MCP Tools area
  2. Click Add MCP Server
  3. Fill in the configuration:
    • Name: Give this MCP server a recognizable name
    • Transport method: Choose stdio / SSE / HTTP
    • For stdio: Enter the command (e.g., npx) and arguments (e.g., @anthropic/mcp-server-filesystem)
    • For SSE/HTTP: Enter the server URL
  4. If authentication is required, select the authentication method and enter the credentials
  5. Click Save
  6. The system automatically connects to the server and discovers the available tools
  7. The discovered tools appear in the tool list and are enabled by default

Tip: MCP subprocesses only pass whitelisted environment variables. If an MCP tool requires environment variables such as an API Key, they must be explicitly declared in the configuration's env field.

Impact of Disabling a Tool

After a tool is disabled, the Agent is completely unaware of that tool's existence in conversations. This means:

Disabled Tool Impact
Bash The Agent cannot execute any shell commands (npm, git, compilation, etc.)
Read File / Write File The Agent cannot read or create files
Web Fetch The Agent cannot fetch web page content, and search-related skills are also affected
Memory Query / Manage The Agent will not query or store memory
Skill The Agent cannot invoke any Skill module
create_local_sub_agent The Agent cannot create sub-agents to break down tasks

Recommendation: Unless there is a clear security reason, we do not recommend disabling the core file operation tools (Read / Write / Edit / Glob / Search), otherwise the Agent's task execution capabilities will be severely limited.

FAQ

The Agent says "I cannot perform file operations"

Check whether the file operation tools (Read File, Write File, Edit, etc.) have been disabled. Confirm on the Tool Management page that these tools are enabled.

MCP Tool Connection Failure

  1. Confirm that the MCP server's command/URL is correct
  2. If using the stdio method, confirm that the corresponding command is installed (e.g., npx requires a Node.js environment)
  3. Check whether environment variables (such as an API Key) are required
  4. Check the APP logs for detailed error information

The Agent Isn't Using the Tool I Expected

The Agent autonomously selects tools based on task requirements. If the Agent chooses the "wrong" tool, you can give explicit instructions in the conversation: "Please use the Bash tool to run this command."