Tools - PicoClaw

Overview

Tools extend agent capabilities by providing interfaces to external systems, APIs, and operations. PicoClaw’s tool system is designed for safety, extensibility, and minimal overhead.

Tool Interface

All tools implement the Tool interface (pkg/tools/types.go):

type Tool interface {
    Name() string                                    // Tool identifier
    Description() string                              // Human-readable description
    Parameters() map[string]any                       // JSON Schema for arguments
    Execute(ctx context.Context, args map[string]any) *ToolResult
}

type ToolResult struct {
    ForLLM  string   // Result sent to LLM
    ForUser string   // Result shown to user (if different)
    Silent  bool     // If true, don't send ForUser to user
    IsError bool     // Indicates execution failure
    Err     error    // Original error (if any)
    Async   bool     // Tool running in background
    Media   []string // Media references (media://...)
}

Optional Interfaces

ContextualTool

Tools that need channel/chat context:

type ContextualTool interface {
    Tool
    SetContext(channel, chatID string)
}

Examples: message, spawn, cron

AsyncTool

Tools that run in background:

type AsyncTool interface {
    Tool
    SetCallback(cb AsyncCallback)
}

type AsyncCallback func(ctx context.Context, result *ToolResult)

Examples: spawn

Tool Registry

Central registry for all available tools (pkg/tools/registry.go):

type ToolRegistry struct {
    tools map[string]Tool
    mu    sync.RWMutex
}

// Core methods
func (r *ToolRegistry) Register(tool Tool)
func (r *ToolRegistry) Get(name string) (Tool, bool)
func (r *ToolRegistry) Execute(ctx context.Context, name string, args map[string]any) *ToolResult
func (r *ToolRegistry) ToProviderDefs() []providers.ToolDefinition

Features:

Thread-safe registration and execution
Automatic conversion to LLM tool definitions
Deterministic ordering (sorted alphabetically) for cache stability
Context injection for contextual tools
Async callback support

Available Tools

Filesystem Tools

read_file

Read file contents. Parameters:

{
  "path": "./data/config.json"
}

Restrictions:

Restricted to workspace if restrict_to_workspace: true
Respects allow_read_paths patterns
Symlink escape prevention

Implementation: pkg/tools/filesystem.go

write_file

Write content to file. Parameters:

{
  "path": "./output.txt",
  "content": "Hello, world!"
}

Features:

Atomic write with sync (flash-safe)
Creates parent directories automatically
Uses secure permissions (0o600)

Restrictions:

Restricted to workspace if restrict_to_workspace: true
Respects allow_write_paths patterns

list_dir

List directory contents. Parameters:

{
  "path": "./src"
}

Returns:

DIR:  components
FILE: main.go
FILE: README.md

edit_file

Search and replace in files. Parameters:

{
  "path": "./config.json",
  "old_text": "\"debug\": false",
  "new_text": "\"debug\": true"
}

Implementation: pkg/tools/edit.go

append_file

Append content to file. Parameters:

{
  "path": "./log.txt",
  "content": "New log entry\n"
}

Shell Tools

exec

Execute shell commands with safety guards. Parameters:

{
  "command": "ls -la",
  "working_dir": "./src"  // Optional
}

Safety Features:

Dangerous Pattern Blocking:
- rm -rf, del /f, rmdir /s
- format, mkfs, diskpart
- dd if=, writes to /dev/sd*
- shutdown, reboot, poweroff
- Fork bombs
- Command substitution ($(...), `...`)
- Pipe to shell (| sh, | bash)
- sudo, chmod, chown
- Package managers (apt, yum, npm -g)
- docker run/exec
- git push/force
- eval, source
Workspace Restriction:
- Absolute paths validated against workspace
- Path traversal blocked (../)
- Safe paths allowed (/dev/null, /dev/urandom, etc.)

Custom Patterns:

{
  "tools": {
    "exec": {
      "enable_deny_patterns": true,
      "custom_deny_patterns": ["\\bcurl\\b.*--upload-file"],
      "custom_allow_patterns": ["^ls\\s"]
    }
  }
}

Timeout: Default 60s, configurable

Implementation: pkg/tools/shell.go Security Notes:

Even with restrict_to_workspace: false, dangerous commands are blocked
Use custom_allow_patterns to override specific blocks

Web Tools

web_search

Search the web for information. Parameters:

{
  "query": "latest AI news",
  "count": 5  // Optional, 1-10
}

Supported Providers (priority order):

Perplexity (LLM-enhanced search)
Brave Search (fast, high-quality)
Tavily (AI-optimized)
DuckDuckGo (free, no API key)

Configuration:

{
  "tools": {
    "web": {
      "brave": {
        "enabled": true,
        "api_key": "BSA...",
        "max_results": 5
      },
      "duckduckgo": {
        "enabled": true,
        "max_results": 5
      }
    }
  }
}

Returns:

Results for: latest AI news
1. AI Breakthrough Announced
   https://example.com/ai-news
   Description of the article...

2. New Model Released
   https://example.com/model
   Details about the model...

Implementation: pkg/tools/web.go

web_fetch

Fetch and extract content from URLs. Parameters:

{
  "url": "https://example.com/article",
  "maxChars": 10000  // Optional
}

Features:

HTML to text conversion
JSON formatting
Redirect following (max 5)
Size limit (default 10MB)
Timeout: 60s

Returns:

{
  "url": "https://example.com/article",
  "status": 200,
  "extractor": "text",
  "truncated": false,
  "length": 5432,
  "text": "Article content..."
}

Communication Tools

message

Send message to user on a chat channel. Parameters:

{
  "content": "Task completed successfully!",
  "channel": "telegram",  // Optional, uses current channel
  "chat_id": "123456"     // Optional, uses current chat
}

Use Cases:

Subagent → User communication
Multi-channel messaging
Background task notifications

Behavior:

Returns Silent: true (user already received message)
Tracks sends per round to prevent duplicates

Implementation: pkg/tools/message.go

Scheduling Tools

cron

Schedule reminders, tasks, or commands. Actions:

add

Create new scheduled task. One-time reminder:

{
  "action": "add",
  "message": "Meeting in 10 minutes",
  "at_seconds": 600  // 10 minutes from now
}

Recurring task:

{
  "action": "add",
  "message": "Daily standup reminder",
  "every_seconds": 86400,  // Daily
  "deliver": true  // Send directly to chat
}

Cron expression:

{
  "action": "add",
  "message": "Check server status",
  "cron_expr": "0 9 * * *",  // Daily at 9am
  "deliver": false  // Process through agent
}

Shell command:

{
  "action": "add",
  "message": "Check disk usage",
  "command": "df -h",
  "every_seconds": 3600  // Hourly
}

list

List all scheduled jobs.

{"action": "list"}

remove

Delete a job.

{
  "action": "remove",
  "job_id": "abc123"
}

enable/disable

Toggle job execution.

{
  "action": "disable",
  "job_id": "abc123"
}

Job Execution:

deliver: true: Send message directly to channel
deliver: false: Process through agent (for complex tasks)
command: Execute shell command, report output

Implementation: pkg/tools/cron.go

Agent Spawning Tools

spawn

Spawn subagent for background tasks. Parameters:

{
  "task": "Search for AI news and write a summary report",
  "label": "AI News Research",  // Optional, for display
  "agent_id": "researcher"       // Optional, target agent
}

Behavior:

Creates independent subagent with fresh context
Subagent has access to all tools (message, web, etc.)
Non-blocking (main agent continues)
Subagent communicates via message tool

Permissions: Controlled via subagents.allowlist config:

{
  "agents": {
    "agents": [
      {
        "id": "main",
        "subagents": {
          "allowlist": ["researcher", "coder"]
        }
      }
    ]
  }
}

Use Cases:

Long-running research tasks
Web scraping and analysis
Background monitoring
Parallel task execution

Implementation: pkg/tools/spawn.go

Hardware Tools (Linux only)

i2c

I2C bus communication. Parameters:

{
  "bus": 1,
  "address": 0x48,
  "operation": "read",
  "register": 0x00,
  "length": 2
}

Availability: Linux only, returns error on other platforms Implementation: pkg/tools/i2c.go

spi

SPI bus communication. Parameters:

{
  "device": "/dev/spidev0.0",
  "data": [0x01, 0x02, 0x03]
}

Availability: Linux only Implementation: pkg/tools/spi.go

Skill Management Tools

find_skills

Search skill registries. Parameters:

{
  "query": "web scraping"
}

Implementation: pkg/tools/skills_search.go

install_skill

Install skill from registry. Parameters:

{
  "skill_id": "web-scraper",
  "registry": "clawhub"  // Optional
}

Implementation: pkg/tools/skills_install.go

MCP Tools

Dynamic tools loaded from MCP (Model Context Protocol) servers. Configuration:

{
  "tools": {
    "mcp": {
      "enabled": true,
      "servers": {
        "filesystem": {
          "command": "npx",
          "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
        }
      }
    }
  }
}

Features:

Auto-registers tools from connected MCP servers
Tool names prefixed with server name (e.g., filesystem_read)
Supports all MCP tool types

Implementation: pkg/tools/mcp_tool.go

Tool Execution Flow

1. LLM returns tool call
   ↓
2. AgentLoop extracts tool calls
   ↓
3. For each tool call:
   │
   ├─ Lookup tool in registry
   │
   ├─ If ContextualTool: SetContext(channel, chatID)
   │
   ├─ If AsyncTool: SetCallback(asyncCallback)
   │
   ├─ Execute tool with args
   │
   ├─ If ForUser and not Silent: Send to user
   │
   ├─ If Media: Send media to user
   │
   └─ Append ForLLM to messages
   ↓
4. Continue to next LLM iteration

Tool Safety

Workspace Sandboxing

When restrict_to_workspace: true (default): File Tools:

All paths resolved relative to workspace
Absolute paths validated
Symlink escape prevention
Path traversal blocked

Shell Tool:

Absolute paths in commands validated
Workspace boundary enforced
Dangerous commands blocked

Example:

# Workspace: /home/user/.picoclaw/workspace

# Allowed:
./data/file.txt          → /home/user/.picoclaw/workspace/data/file.txt
/home/user/.picoclaw/workspace/file.txt → OK

# Blocked:
../../../etc/passwd      → Access denied (path traversal)
/etc/passwd              → Access denied (outside workspace)
/home/user/documents/    → Access denied (outside workspace)

Path Whitelisting

Allow specific paths outside workspace:

{
  "tools": {
    "allow_read_paths": [
      "^/etc/hosts$",
      "^/proc/cpuinfo$"
    ],
    "allow_write_paths": [
      "^/tmp/picoclaw-.*"
    ]
  }
}

Pattern Syntax: Regular expressions

Exec Safety Guards

Multi-layer protection for shell commands:

Deny Patterns: Block dangerous commands
Allow Patterns: Override blocks for specific cases
Workspace Restriction: Path validation
Timeout: Prevent infinite execution

Disable All Deny Patterns (dangerous):

{
  "tools": {
    "exec": {
      "enable_deny_patterns": false
    }
  }
}

Tool Result Patterns

Standard Result

return &ToolResult{
    ForLLM:  "File written successfully",
    ForUser: "Saved to output.txt",
}

Silent Result

return &ToolResult{
    ForLLM:  "Message sent",
    Silent:  true,  // Don't send ForUser to user
}

Error Result

return &ToolResult{
    ForLLM:  "Failed to read file: file not found",
    IsError: true,
    Err:     err,
}

Async Result

return &ToolResult{
    ForLLM: "Subagent spawned, task running in background",
    Async:  true,
}

Media Result

return &ToolResult{
    ForLLM: "Screenshot captured",
    Media:  []string{"media://abc123"},
}

Creating Custom Tools

1. Implement Tool Interface

package customtools

import (
    "context"
    "github.com/sipeed/picoclaw/pkg/tools"
)

type GreetTool struct{}

func (t *GreetTool) Name() string {
    return "greet"
}

func (t *GreetTool) Description() string {
    return "Greet a person by name"
}

func (t *GreetTool) Parameters() map[string]any {
    return map[string]any{
        "type": "object",
        "properties": map[string]any{
            "name": map[string]any{
                "type":        "string",
                "description": "Person's name",
            },
        },
        "required": []string{"name"},
    }
}

func (t *GreetTool) Execute(ctx context.Context, args map[string]any) *tools.ToolResult {
    name, ok := args["name"].(string)
    if !ok {
        return tools.ErrorResult("name is required")
    }
    
    greeting := fmt.Sprintf("Hello, %s!", name)
    return tools.NewToolResult(greeting)
}

2. Register Tool

Modify pkg/agent/loop.go:

func registerSharedTools(...) {
    for _, agentID := range registry.ListAgentIDs() {
        agent, _ := registry.GetAgent(agentID)
        
        // ... existing tools ...
        
        // Register custom tool
        agent.Tools.Register(&customtools.GreetTool{})
    }
}

3. Rebuild and Test

make build
./build/picoclaw agent -m "Greet Alice"

Best Practices

1. Tool Naming

Use lowercase with underscores: read_file, web_search
Be descriptive and action-oriented
Avoid conflicts with existing tools

2. Parameter Validation

Always validate required parameters
Provide clear error messages
Use JSON Schema for type safety

3. Error Handling

Return ErrorResult() for user-facing errors
Include original error in Err field
Keep error messages concise and actionable

4. Performance

Set appropriate timeouts
Avoid blocking operations in Execute()
Use Async pattern for long tasks

5. Security

Validate all user inputs
Sanitize file paths
Use workspace restrictions
Follow principle of least privilege

Get Started

Core Concepts

Commands

Configuration

Chat Channels

Features

Deployment

Advanced

​Overview

​Tool Interface

​Optional Interfaces

​ContextualTool

​AsyncTool

​Tool Registry

​Available Tools

​Filesystem Tools

​read_file

​write_file

​list_dir

​edit_file

​append_file

​Shell Tools

​exec

​Web Tools

​web_search

​web_fetch

​Communication Tools

​message

​Scheduling Tools

​cron

add

list

remove

enable/disable

​Agent Spawning Tools

​spawn

​Hardware Tools (Linux only)

​i2c

​spi

​Skill Management Tools

​find_skills

​install_skill

​MCP Tools

​Tool Execution Flow

​Tool Safety

​Workspace Sandboxing

​Path Whitelisting

​Exec Safety Guards

​Tool Result Patterns

​Standard Result

​Silent Result

​Error Result

​Async Result

​Media Result

​Creating Custom Tools

​1. Implement Tool Interface

​2. Register Tool

​3. Rebuild and Test

​Best Practices

​1. Tool Naming

​2. Parameter Validation

​3. Error Handling

​4. Performance

​5. Security

Build docs developers (and LLMs) love

Overview

Tool Interface

Optional Interfaces

ContextualTool

AsyncTool

Tool Registry

Available Tools

Filesystem Tools

read_file

write_file

list_dir

edit_file

append_file

Shell Tools

exec

Web Tools

web_search

web_fetch

Communication Tools

message

Scheduling Tools

cron

Agent Spawning Tools

spawn

Hardware Tools (Linux only)

i2c

spi

Skill Management Tools

find_skills

install_skill

MCP Tools

Tool Execution Flow

Tool Safety

Workspace Sandboxing

Path Whitelisting

Exec Safety Guards

Tool Result Patterns

Standard Result

Silent Result

Error Result

Async Result

Media Result

Creating Custom Tools

1. Implement Tool Interface

2. Register Tool

3. Rebuild and Test

Best Practices

1. Tool Naming

2. Parameter Validation

3. Error Handling

4. Performance

5. Security