ChannelCoder

A streamlined SDK and CLI for Claude Code that channels your prompts with powerful features like multi-place variable interpolation, file-based prompts, and streaming support.

Installation

# Using npm
npm install channelcoder

# Using bun
bun add channelcoder

# Global CLI installation
npm install -g channelcoder

Quick Start

SDK Usage

import { claude } from 'channelcoder';

// Simple prompt
const result = await claude('What is TypeScript?');

// Template literals
const language = 'TypeScript';
await claude`Explain ${language} in simple terms`;

// File-based prompt with data
await claude('prompts/analyze.md', {
  data: { taskId: 'FEAT-123', priority: 'high' }
});

// With options
await claude('Review this code', {
  tools: ['Read', 'Grep'],
  system: 'You are a code reviewer'
});

CLI Usage

The CLI provides multiple commands for different execution modes:

# Interactive mode (default) - run Claude interactively
channelcoder interactive prompts/analyze.md -d taskId=FEAT-123

# Run mode - execute and print results
channelcoder run "Explain TypeScript"
channelcoder run prompts/analyze.md --data taskId=FEAT-123

# Stream mode - real-time streaming responses  
channelcoder stream "Write a story"
channelcoder stream prompts/generate.md

# Session management
channelcoder session list                    # List saved sessions
channelcoder session load my-session         # Load and continue a session
channelcoder session remove old-session      # Remove a session

# Worktree management
channelcoder worktree list                   # List git worktrees
channelcoder worktree create feature/new     # Create new worktree
channelcoder worktree remove feature/old     # Remove worktree

Features

🎯 Simple Function API

Mirrors Claude CLI's mental model with a simple function:

// Inline prompts
await claude('Explain quantum computing');

// File prompts (auto-detected by .md extension)
await claude('prompts/complex-analysis.md');

// With all the options you need
await claude('Debug this issue', {
  tools: ['Read', 'Bash'],
  resume: sessionId,
  maxTurns: 10
});

🔄 Variable Interpolation

Powerful multi-place variable support:

// In inline prompts
await claude('Analyze {code} for {issues}', {
  data: {
    code: 'const x = null',
    issues: ['null safety', 'type errors']
  }
});

// In file prompts
// prompts/review.md:
// Review {prTitle} with focus on {concerns ? concerns : "general quality"}
await claude('prompts/review.md', {
  data: { prTitle: 'Add auth system' }
});

🌊 Execution Modes

Different modes for different needs:

import { claude, interactive, stream, run, detached } from 'channelcoder';

// Run mode (default) - Get complete results programmatically
const result = await claude('Generate tests');
console.log(result.data);

// Stream mode - Real-time responses
for await (const chunk of stream('Write documentation')) {
  process.stdout.write(chunk.content);
}

// Interactive mode - Replace process (like running claude directly)
await interactive('Debug this error');
// ⚠️ Code after interactive() never executes!

// Detached mode - Background execution
const result3 = await detached('Long analysis task', {
  logFile: 'analysis.log'
});
console.log('Started background process:', result3.data?.pid);

// Detached streaming - Real-time background monitoring
const result4 = await detached('Generate comprehensive report', {
  logFile: 'report.log',
  stream: true  // Enables real-time JSON chunks in log file
});
// Monitor with: tail -f report.log | jq -r '.content'

🌿 Git Worktree Mode

Run Claude in isolated git worktrees for parallel development and clean experimentation:

// Simple worktree usage - auto-creates if doesn't exist
await claude('Implement new feature', { 
  worktree: 'feature/auth' 
});

// Standalone worktree function for complex workflows
await worktree('feature/payments', async (wt) => {
  console.log(`Working in: ${wt.path}`);
  await claude('Design payment system');
  await claude('Implement payment processing');
  return 'feature-complete';
}, {
  base: 'main',      // Create from main branch
  cleanup: false     // Keep worktree after execution
});

// Worktree utilities for advanced management
import { worktreeUtils } from 'channelcoder';

// List all worktrees
const worktrees = await worktreeUtils.list();

// Create/remove worktrees manually
await worktreeUtils.create('experiment/new-arch', { base: 'develop' });
await worktreeUtils.remove('feature/old');

// Check worktree existence
if (await worktreeUtils.exists('feature/auth')) {
  console.log('Auth worktree already exists');
}

Why Use Worktrees?

1. Parallel Development: Work on multiple features without stashing/switching
2. Clean Experiments: Test risky changes in isolated environments
3. Context Preservation: Each worktree maintains its own file state
4. Session Enhancement: Combine with sessions for branch-specific conversations

Worktree Options

interface WorktreeOptions {
  // Branch name (required for creation)
  branch?: string;
  
  // Base branch for new worktree
  base?: string;
  
  // Custom path (auto-generated if not provided)
  path?: string;
  
  // Remove worktree after execution (default: true for temporary)
  cleanup?: boolean;
  
  // Force creation vs. error if exists (default: false - upsert)
  create?: boolean;
}

Composing Features

Worktrees work seamlessly with other ChannelCoder features:

// Worktree + Session
const s = session();
await s.claude('Start OAuth feature', { worktree: 'feature/oauth' });
await s.claude('Add tests'); // Continues in same worktree

// Worktree + Docker
await claude('Test in total isolation', {
  worktree: 'experiment/risky',
  docker: true
});

// Worktree + Streaming
await worktree('feature/docs', async () => {
  for await (const chunk of stream('Generate documentation')) {
    process.stdout.write(chunk.content);
  }
});

🐳 Docker Mode

Run Claude in an isolated Docker container for enhanced security with dangerous permissions:

// Simple Docker mode - auto-detect Dockerfile
await claude('Risky operation', { docker: true });

// Specify Docker image
await claude('Analyze system', { 
  docker: { image: 'my-claude:latest' } 
});

// With custom mounts and environment
await claude('Process data', {
  docker: {
    image: 'claude-sandbox',
    mounts: ['./data:/data:ro'],
    env: { NODE_ENV: 'production' }
  }
});

Setting Up Docker Mode

1. Create a Dockerfile with Claude CLI installed:

FROM node:20-slim
RUN npm install -g @anthropic-ai/claude-code
WORKDIR /workspace

2. Build your image:

docker build -t my-claude .

3. Use with channelcoder:

await claude('Task', { docker: { image: 'my-claude' } });

Auto-Detection

When using docker: true, channelcoder will:

1. Look for a Dockerfile in your project
2. Build an image automatically (with caching)
3. Mount your working directory as /workspace

Docker Options

interface DockerOptions {
  // Auto-detect Dockerfile (default: true)
  auto?: boolean;
  
  // Docker image to use
  image?: string;
  
  // Path to Dockerfile (default: ./Dockerfile)
  dockerfile?: string;
  
  // Additional volume mounts
  mounts?: string[];  // Docker format: "host:container:mode"
  
  // Environment variables
  env?: Record<string, string>;
}

CLI Reference

channelcoder <command> [options]

Commands:
  run          Execute prompt and exit (print mode) - non-interactive
  interactive  Interactive mode with Claude (default)
  stream       Stream responses in real-time
  session      Manage conversation sessions
  worktree     Manage git worktrees

Global Options:
  -h, --help     Show help
  -v, --version  Show version

Run 'channelcoder <command> --help' for command-specific options

Command Details

channelcoder run [prompt-file] [options]

Execute a prompt and print the result without interaction.

Options:
  -p, --prompt <text>      Inline prompt instead of file
  -d, --data <key=value>   Data for interpolation (repeatable)
  -s, --system <prompt>    System prompt (text or .md file)
  -t, --tools <tools>      Allowed tools (e.g., "Read Write")
  --disallowed-tools       Disallowed tools (comma-separated)
  --append-system <text>   Append to system prompt
  --mcp-config <file>      Load MCP servers from JSON file
  --permission-tool <tool> MCP tool for permission prompts
  -r, --resume <id>        Resume conversation by session ID
  -c, --continue           Continue most recent conversation
  --max-turns <n>          Limit agentic turns
  --json                   Output JSON format
  -v, --verbose            Verbose output

channelcoder interactive [prompt-file] [options]

Launch Claude in interactive mode (default command).

Options: Same as 'run' command

channelcoder stream [prompt-file] [options]

Stream responses in real-time.

Options: Same as 'run' command

channelcoder session <subcommand>

Manage conversation sessions.

Subcommands:
  list       List all saved sessions
  load       Load and continue a session
  remove     Remove a saved session

channelcoder worktree <subcommand>

Manage git worktrees for isolated development.

Subcommands:
  list       List all worktrees
  create     Create a new worktree
  remove     Remove a worktree
  cleanup    Clean up orphaned worktrees

Advanced Features Documentation

For detailed guides on advanced features, see the /documentation folder:

• Docker Mode Guide - Run Claude in isolated containers with enhanced security
• Git Worktree Mode Guide - Parallel development in isolated git worktrees
• Session Management Guide - Multi-turn conversations with context preservation
• Stream Parser SDK Guide - Parse and monitor Claude's streaming output

Frontmatter Syntax

File-based prompts support YAML frontmatter for configuration. Here are the actually supported options:

---
# System prompt - Sets the Claude system/assistant prompt
# Can be inline text or a path to a .md/.txt file
systemPrompt: "You are a helpful coding assistant"
# or
systemPrompt: "./system-prompts/analyst.md"

# Append to system prompt
appendSystemPrompt: "Always explain your reasoning"

# Allowed tools - Restrict which tools Claude can use
# These are passed to Claude CLI's --allowedTools flag
allowedTools:
  - "Read"                    # Read files
  - "Write"                   # Write files
  - "Edit"                    # Edit files
  - "Bash"                    # Run any bash command
  - "Bash(git:*)"            # Pattern: only git commands
  - "Bash(npm:test)"         # Specific: only npm test
  - "Grep"                   # Search file contents
  - "WebSearch"              # Search the web

# Disallowed tools - Prevent specific tools
disallowedTools:
  - "Bash(rm:*)"             # No rm commands
  - "Bash(git:push)"         # No git push

# MCP (Model Context Protocol) configuration
mcpConfig: "./mcp-servers.json"
permissionPromptTool: "mcp__auth__prompt"

# Input schema - Validates variables before interpolation
input:
  name: string               # Required string
  age?: number              # Optional number
  tags: string[]            # Array of strings
  config:                   # Nested object
    port: number
    host?: string

# Output schema - Validates Claude's response (SDK only)
output:
  success: boolean
  result:
    type: string
    enum: [feature, bug, chore]  # Enum constraint
  items: 
    - name: string
      done: boolean
---
Your prompt content here...

Note: The following options are passed via CLI or SDK, not frontmatter:

• outputFormat - Use --json flag or claude(prompt, { outputFormat: 'json' })
• stream - Use --stream flag or stream() function
• verbose - Use --verbose flag or claude(prompt, { verbose: true })
• timeout - SDK only: claude(prompt, { timeout: 30000 })

Frontmatter Validation

The frontmatter is validated using a Zod schema. Invalid keys will cause an error:

import { FrontmatterSchema, type Frontmatter } from 'channelcoder';

// Validate frontmatter programmatically
const result = FrontmatterSchema.safeParse({
  systemPrompt: "Valid",
  temperature: 0.7  // Error: unknown key
});

// TypeScript type for frontmatter
const config: Frontmatter = {
  systemPrompt: "Assistant prompt",
  appendSystemPrompt: "Be concise",
  allowedTools: ["Read", "Write"],
  disallowedTools: ["Bash(rm:*)"],
  mcpConfig: "./mcp-servers.json",
  permissionPromptTool: "mcp__auth__prompt",
  input: { name: "string" },
  output: { success: "boolean" }
};

Schema Definition Format

Schemas are defined using YAML notation that's automatically converted to Zod schemas:

---
input:
  name: string              # Basic types: string, number, boolean
  age?: number             # Optional with ?
  tags: string[]           # Arrays with []
  metadata:                # Nested objects
    created: string
    updated?: string
---

Supported types:

• string - Text values
• number - Numeric values
• boolean or bool - True/false values
• array or type[] - Arrays
• object - Nested objects
• any - Any value

Note: For programmatic SDK usage, you can also use Zod schemas directly in the validation utilities.

Examples

File-based Prompts

Create a prompt file analyze.md:

---
input:
  task: string
  details?: boolean
systemPrompt: "You are a helpful analyst"
allowedTools:
  - Read
  - Grep
---
# Analysis for {task}

{details ? "Provide detailed breakdown." : "Summary only."}

Run it:

channelcoder analyze.md -d task="Review PR" -d details=true

Inline Prompts

# Simple
channelcoder -p "Explain {concept}" -d concept="quantum computing"

# With tools
channelcoder -p "Find files containing {pattern}" \
   -d pattern="TODO" \
   -t "Read Grep"

# Streaming
channelcoder -p "Write a haiku about {topic}" \
   -d topic="coding" \
   --stream

# Resume conversation
channelcoder --resume abc123 -p "Continue with the implementation"

# Continue last conversation
channelcoder --continue -p "What about error handling?"

# Limit agentic turns
channelcoder analyze.md --max-turns 3

# MCP configuration
channelcoder query.md --mcp-config ./servers.json

# Disallow dangerous tools
channelcoder cleanup.md --disallowed-tools "Bash(rm:*),Bash(git:push)"

Complex Data

# Arrays and objects via JSON
channelcoder -p "Process items: {items}" \
   -d 'items=["apple","banana","orange"]'

# Complex nested data
channelcoder -p "Config: {config}" \
   -d 'config={"port":3000,"host":"localhost"}'

SDK Reference

Basic Usage

import { claude } from 'channelcoder';

// Simple prompts
const result = await claude('Explain TypeScript');

// Template literals
const topic = 'async/await';
const result = await claude`Explain ${topic} with examples`;

// File-based prompts
const result = await claude('prompts/analyze.md', {
  data: { taskId: 'FEAT-123' }
});

Options

// All available options
const result = await claude('Your prompt', {
  // Data interpolation
  data: { key: 'value' },
  
  // System configuration
  system: 'You are a helpful assistant',
  appendSystem: 'Be concise',
  
  // Tool configuration
  tools: ['Read', 'Write', 'Bash(git:*)'],
  disallowedTools: ['Bash(rm:*)'],
  mcpConfig: './mcp-servers.json',
  permissionTool: 'mcp__auth__prompt',
  
  // Session management
  resume: 'session-id-here',
  continue: true,
  
  // Execution control
  maxTurns: 10,
  mode: 'run', // 'run' | 'stream' | 'interactive'
  includeEvents: true,
  
  // Background execution
  detached: true,        // Run in background
  logFile: 'output.log', // Log file for detached mode
  stream: true,          // Enable real-time streaming in detached mode
  
  // Other
  verbose: true,
  outputFormat: 'json',
  timeout: 60000
});

Streaming

import { stream } from 'channelcoder';

// Stream responses
for await (const chunk of stream('Generate a story')) {
  if (chunk.type === 'content') {
    process.stdout.write(chunk.content);
  }
}

Error Handling

const result = await claude('prompt.md', { data });

if (!result.success) {
  console.error('Error:', result.error);
  if (result.warnings) {
    console.warn('Warnings:', result.warnings);
  }
}

// Type-safe error handling
if (result.warnings?.includes('max turns reached')) {
  console.log('Hit maximum turn limit');
}

Interactive Mode

import { interactive } from 'channelcoder';

// Launch Claude interactively (replaces current process)
await interactive('Help me debug this issue');

// Code below this line will NEVER execute!
// The process is replaced by Claude using shell exec

Important: Interactive mode completely replaces your Node.js process with Claude:

• No Node.js parent process remains in memory
• Claude gets direct terminal control
• Exit codes go directly to the shell
• Perfect for long Claude sessions without memory overhead

Background Execution & Real-time Monitoring

ChannelCoder supports background execution with real-time monitoring capabilities:

import { detached, session } from 'channelcoder';

// Basic background execution
const result = await detached('Analyze large codebase', {
  logFile: 'analysis.log'
});
console.log('Background process started with PID:', result.data?.pid);

// Real-time streaming to log file
const result2 = await detached('Generate comprehensive report', {
  logFile: 'report.log',
  stream: true  // Enables real-time JSON streaming
});

// Monitor in real-time with Unix tools
// tail -f report.log | jq -r '.content'          # See content
// tail -f report.log | jq -r '.type'             # See chunk types
// watch -n 1 "wc -l report.log"                  # Monitor progress

Session Background Execution:

// Session with real-time file updates
const s = session({ 
  autoSave: true  // Session file updates in real-time
});

// Background execution with session context
await s.detached('Long-running analysis', {
  logFile: 'session-output.log',
  stream: true
});

// Monitor both Claude output AND session state
// tail -f session-output.log | jq -r '.content'   # Claude output
// watch -n 1 cat ~/.channelcoder/sessions/my-session.json  # Session state

Stream Parser SDK

ChannelCoder includes a powerful Stream Parser SDK for parsing Claude's stream-json output from detached sessions and log files:

import { parseLogFile, monitorLog, streamParser } from 'channelcoder';

// Parse a complete log file
const parsed = await parseLogFile('session.log');
console.log(parsed.content);        // All assistant messages
console.log(parsed.totalCost);      // Total cost
console.log(parsed.events.length);  // Number of events

// Monitor a log file in real-time
const cleanup = monitorLog('active.log', (event) => {
  if (streamParser.isAssistantEvent(event)) {
    console.log('Claude:', streamParser.extractAssistantText(event));
  } else if (streamParser.isToolUseEvent(event)) {
    console.log('Tool used:', event.tool);
  } else if (streamParser.isResultEvent(event)) {
    console.log('Completed! Cost:', event.cost_usd);
  }
});

// Clean up when done
cleanup();

// Low-level parsing
import { parseStreamEvent, eventToChunk } from 'channelcoder/streamParser';

const line = '{"type":"assistant","message":{...}}';
const event = parseStreamEvent(line);
const chunk = eventToChunk(event);

Type Guards for Event Handling:

import { 
  isSystemEvent, 
  isAssistantEvent, 
  isResultEvent,
  isToolUseEvent,
  isErrorEvent 
} from 'channelcoder';

// Type-safe event processing
function processEvent(event: ClaudeEvent) {
  if (isAssistantEvent(event)) {
    const text = streamParser.extractAssistantText(event);
    console.log('Assistant:', text);
  } else if (isToolUseEvent(event)) {
    console.log('Tool:', event.tool, event.input);
  } else if (isResultEvent(event)) {
    if (event.subtype === 'error') {
      console.error('Failed:', event.error);
    } else {
      console.log('Success! Cost: $', event.cost_usd);
    }
  }
}

See examples/task-monitor-tui.ts for a complete real-time monitoring TUI built with the Stream Parser SDK.

Session Management

ChannelCoder provides built-in session management for maintaining conversation context:

import { session } from 'channelcoder';

// Create a new session with auto-save
const s = session({ autoSave: true });

// Use like normal, but with automatic context tracking
await s.claude('What is TypeScript?');
await s.claude('Show me an example'); // Automatically continues conversation

// Save session for later
await s.save('learning-typescript');

// Load and continue a saved session
const saved = await session.load('learning-typescript');
await saved.claude('What about generics?');

// List all saved sessions
const sessions = await session.list();
// [{ name: 'learning-typescript', messageCount: 3, lastActive: Date, ... }]

// Access session data
console.log(s.id());         // Current session ID
console.log(s.messages());   // Conversation history

CLI Session Support:

# Start a new session
channelcoder prompts/debug.md --session my-debug

# Continue a session
channelcoder prompts/continue.md --load-session my-debug

# List all sessions
channelcoder --list-sessions

Session-Required Prompts:

---
session:
  required: true
systemPrompt: "You are debugging an ongoing issue"
---
Continue investigating the error we discussed.

Manual Session Management (without session wrapper):

// Resume a specific session by ID
await claude('Continue our discussion', {
  resume: 'session-id-here'
});

// Continue most recent session
await claude('Continue where we left off', {
  continue: true
});

Prompt File Format

ChannelCoder uses Markdown files with YAML frontmatter:

---
# Input validation (optional)
input:
  name: string
  age?: number
  tags: string[]

# Output schema (optional)
output:
  success: boolean
  message: string

# Claude options
systemPrompt: "path/to/system.md"  # or inline text
allowedTools:
  - Read
  - Write
  - "Bash(git:*)"  # With patterns
---

# Your prompt with {name} interpolation

Age: {age ? age : "not specified"}

Tags: {tags}

Requirements

• Claude Code CLI installed and configured
• Node.js 18+ or Bun runtime

Tips & Tricks

1. Tool Patterns

Allow specific command patterns:

channelcoder prompt.md -t "Bash(git:*) Read Write"

2. Conditional Content

Use JavaScript expressions in templates:

const isDev = true;
const items = ['a', 'b', 'c'];

await claude`
  ${isDev ? "Include debug info" : ""}
  Process ${items.length} items
`;

3. System Prompts

Can be inline or file paths:

# Inline
channelcoder -p "..." -s "Be concise"

# From file
channelcoder -p "..." -s "prompts/systems/expert.md"

4. Background Monitoring

Perfect for long-running tasks:

# Start background streaming process
channelcoder analysis.md --detached --stream --log analysis.log

# Monitor real-time progress (Unix composable)
tail -f analysis.log | jq -r '.content'           # Content only
tail -f analysis.log | jq -r 'select(.type=="tool_use").tool'  # Tool usage
watch -n 1 "grep -c content analysis.log"         # Progress counter

5. JSON Output

Perfect for automation:

result=$(cc prompt.md --json)
success=$(echo $result | jq -r '.success')

Examples

Check out the /examples directory for:

• basic-usage.ts - Simple examples to get started
• file-based-usage.ts - Using file-based prompts
• launch-modes.ts - Different execution modes
• detached-streaming.ts - Background execution with real-time monitoring
• task-monitor-tui.ts - NEW Real-time TUI using Stream Parser SDK
• demo-features.ts - Feature showcase (no execution)
• release.ts - Real-world release automation

Run examples:

bun run example:quick     # Run basic examples
bun run examples/basic-usage.ts
bun run examples/launch-modes.ts run
bun run examples/release.ts

License

MIT

Contributing

Issues and PRs welcome at github.com/davidpp/channelcoder

Named after Claude Shannon, the father of information theory, ChannelCoder channels your prompts to Claude with maximum signal and minimum noise.