Imported from reason-machines/codex-skills (skills/seemseam-ccb-multi-agent-cli/SKILL.md). Install upstream with npx skills add reason-machines/codex-skills --skill seemseam-ccb-multi-agent-cli. Copyright stays with the author.

SeemSeam CCB Multi-Agent CLI

Skill by ara.so — Codex Skills collection.

CCB (Claude Codex Bridge) is a multi-agent orchestration framework that runs Claude, Codex, Gemini, OpenCode, and Droid agents in supervised tmux panes with shared project memory, inter-agent communication via /ask, and isolated worktree support for parallel work.

What CCB Does

• Unified CLI entry point: Start, attach, recover, and supervise multiple AI agent CLIs from one command
• Inter-agent communication: Agents can /ask each other, broadcast updates, and delegate work
• Project-level teams: Define role-based teams with custom pane layouts, provider state, and worktree isolation
• Shared memory: All agents access .ccb/ccb_memory.md for project-wide context
• Tmux supervision: Every agent runs in a named tmux pane with lifecycle management

Installation

Unix-like (Linux, macOS, WSL)

git clone https://github.com/SeemSeam/claude_codex_bridge.git
cd claude_codex_bridge
./install.sh install

Windows

git clone https://github.com/SeemSeam/claude_codex_bridge.git
cd claude_codex_bridge
powershell -ExecutionPolicy Bypass -File .\install.ps1 install

Update to Latest Release

ccb update              # Latest stable
ccb update 6            # Highest v6.x.x
ccb update 6.1          # Highest v6.1.x
ccb update 6.1.21       # Specific version

Requirements

• Python 3.10+
• tmux

Core Commands

# Start agents from .ccb/ccb.config
ccb

# Safe start (preserve permission settings)
ccb -s

# Rebuild state (preserve config) then start
ccb -n

# Stop project runtime
ccb kill

# Force cleanup before rebuild
ccb kill -f

# Uninstall
ccb uninstall

# Reinstall
ccb reinstall

Configuration

CCB is configured via .ccb/ccb.config (project-local, user-authored). If missing, CCB uses built-in defaults without creating a file.

Basic Layout Syntax

The first line defines the team and pane layout:

cmd; writer:codex, reviewer:claude; qa:gemini(worktree)

Layout rules:

• ; splits panes left-to-right
• , stacks panes top-to-bottom
• cmd is the shell pane
• name:provider defines an agent
• (worktree) runs agent in isolated git worktree
• Without (worktree), agent runs inplace

Common Layouts

# Two-agent team
writer:codex, reviewer:claude

# Shell + three agents
cmd; writer:codex, reviewer:claude; qa:gemini(worktree)

# Same provider, different roles
cmd; fast:codex, deep:codex

Per-Agent API Configuration

Add TOML tables after the layout line for agents needing custom API keys, URLs, or models:

cmd; builder:codex, reviewer:claude; research:gemini(worktree)

[agents.builder]
key = "$OPENAI_API_KEY"
url = "https://api.openai.com/v1"
model = "gpt-4"

[agents.reviewer]
key = "$ANTHROPIC_API_KEY"
url = "https://api.anthropic.com"
model = "claude-3-5-sonnet-20241022"

[agents.research]
key = "$GEMINI_API_KEY"
model = "gemini-2.0-flash-exp"

Notes:

• Use environment variables for API keys ($VAR_NAME)
• key and url override global provider credentials
• model sets agent-specific model
• Do not commit real API keys

Same Provider, Multiple API Keys

cmd; fast:codex, deep:codex

[agents.fast]
key = "$OPENAI_FAST_KEY"
model = "gpt-4o-mini"

[agents.deep]
key = "$OPENAI_DEEP_KEY"
url = "https://api.example.com/v1"
model = "gpt-4o"

Advanced Provider Environment

[agents.builder.provider_profile.env]
OPENAI_API_KEY = "$OPENAI_BUILDER_KEY"
OPENAI_BASE_URL = "https://custom.endpoint.com/v1"

Do not mix key/url shortcuts with provider_profile.env on the same agent.

Inter-Agent Communication

CCB agents can communicate using /ask or $ask syntax.

Explicit /ask Delegation

/ask reviewer review the parser changes in src/parser.ts

Explicit $ask Delegation

$ask reviewer review the parser changes in src/parser.ts

Implicit Delegation (Natural Language)

Ask reviewer to check the parser edge cases, then summarize the issues back to me.

For implicit delegation to work, add the ask skill basics to your system memory or agent prompt.

Broadcasting to All Agents

Agents can broadcast context updates to all live agents when the whole team needs the same information.

Agent Discovery

Named agents can discover each other and use named targets for delegation without copy/paste.

Project Memory

.ccb/ccb_memory.md is the shared project memory document. All agents in the team can read and write to this file for persistent context.

# Example: Agent updating shared memory
with open('.ccb/ccb_memory.md', 'a') as f:
    f.write('\n## Feature X Implementation\n')
    f.write('- Completed API endpoint `/api/v1/feature`\n')
    f.write('- Added tests in `tests/test_feature.py`\n')

Worktree Isolation

Agents marked with (worktree) run in isolated git worktrees, enabling parallel work without conflicts.

Example: QA Agent in Worktree

cmd; builder:codex, reviewer:claude; qa:gemini(worktree)

The qa agent runs in a separate worktree under .ccb/worktrees/qa/, allowing it to:

• Test changes without affecting main working tree
• Run parallel test suites
• Isolate experimental work

Real-World Examples

Example 1: Full-Stack Development Team

.ccb/ccb.config:

cmd; frontend:codex, backend:claude; test:gemini(worktree)

[agents.frontend]
key = "$OPENAI_API_KEY"
model = "gpt-4o"

[agents.backend]
key = "$ANTHROPIC_API_KEY"
model = "claude-3-5-sonnet-20241022"

[agents.test]
key = "$GEMINI_API_KEY"
model = "gemini-2.0-flash-exp"

Workflow:

1. Start team: ccb
2. Frontend agent builds React component
3. Backend agent implements API endpoint
4. Frontend asks backend: /ask backend does the /api/users endpoint support pagination?
5. Test agent runs integration tests in isolated worktree
6. Test agent reports back: /ask frontend found CORS issue in login flow

Example 2: Code Review Pipeline

.ccb/ccb.config:

cmd; writer:codex, reviewer:claude, qa:codex(worktree)

[agents.writer]
key = "$OPENAI_WRITER_KEY"
model = "gpt-4o"

[agents.reviewer]
key = "$ANTHROPIC_API_KEY"
model = "claude-3-5-sonnet-20241022"

[agents.qa]
key = "$OPENAI_QA_KEY"
model = "gpt-4o-mini"

Workflow:

1. Writer implements feature in src/feature.py
2. Writer asks reviewer: /ask reviewer review src/feature.py for security issues
3. Reviewer provides feedback in chat
4. Writer applies fixes
5. QA agent runs tests in worktree: /ask qa run test suite for feature.py

Example 3: Research and Documentation

cmd; research:gemini, writer:codex

[agents.research]
key = "$GEMINI_API_KEY"
model = "gemini-2.0-flash-exp"

[agents.writer]
key = "$OPENAI_API_KEY"
model = "gpt-4o"

Workflow:

1. Research agent explores API documentation
2. Research broadcasts findings: agent updates .ccb/ccb_memory.md
3. Writer reads memory and generates documentation
4. Writer asks research: /ask research verify these GraphQL schema examples

Python Integration Examples

Programmatically Reading Project Memory

import os

def read_project_memory():
    """Read shared project memory for context."""
    memory_path = os.path.join('.ccb', 'ccb_memory.md')
    if os.path.exists(memory_path):
        with open(memory_path, 'r') as f:
            return f.read()
    return ""

# Use in agent script
context = read_project_memory()
print(f"Current project context:\n{context}")

Writing to Project Memory

import os
from datetime import datetime

def append_to_memory(section, content):
    """Append structured content to project memory."""
    memory_path = os.path.join('.ccb', 'ccb_memory.md')
    timestamp = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
    
    with open(memory_path, 'a') as f:
        f.write(f'\n## {section} ({timestamp})\n\n')
        f.write(content)
        f.write('\n')

# Example usage
append_to_memory(
    'API Endpoint Implementation',
    '- Created `/api/v1/users` endpoint\n'
    '- Added authentication middleware\n'
    '- Tests passing in `tests/test_users.py`'
)

Inter-Agent Ask Wrapper

import subprocess

def ask_agent(agent_name, query):
    """Send query to another agent via CCB ask."""
    result = subprocess.run(
        ['ask', agent_name, query],
        capture_output=True,
        text=True
    )
    return result.stdout

# Example usage
response = ask_agent('reviewer', 'review src/auth.py for security issues')
print(f"Reviewer feedback:\n{response}")

Tmux Integration

CCB runs all agents in tmux panes. Useful tmux commands:

# List CCB sessions
tmux ls

# Attach to CCB session
tmux attach -t <session-name>

# Navigate panes (within tmux)
Ctrl+b <arrow-key>

# Copy mode
# Drag left mouse button to select, Ctrl+Shift+V to paste

Troubleshooting

Issue: ccb command not found

Solution:

# Verify installation
which ccb

# Reinstall
cd claude_codex_bridge
./install.sh install

# Check PATH includes CCB bin directory
echo $PATH | grep ccb

Issue: Agents not starting

Solution:

# Check .ccb/ccb.config syntax
cat .ccb/ccb.config

# Rebuild state
ccb kill -f
ccb -n

# Check agent provider availability
which claude
which codex

Issue: /ask not working

Causes:

• Agent doesn't have ask skill in system memory
• Agent is using built-in multi-agent behavior instead

Solution: Add to agent system prompt or .ccb/ccb_memory.md:

## Inter-Agent Communication

Use `/ask <agent_name> <query>` to delegate tasks to other agents.
Use `$ask <agent_name> <query>` as alternative syntax.

Available agents: [list agent names from layout]

Issue: API key errors

Solution:

# Verify environment variables
echo $OPENAI_API_KEY
echo $ANTHROPIC_API_KEY

# Check .ccb/ccb.config uses env vars
cat .ccb/ccb.config

# Never commit real keys
git diff .ccb/ccb.config

Issue: Worktree conflicts

Solution:

# List worktrees
git worktree list

# Remove stale worktree
git worktree remove .ccb/worktrees/<agent-name>

# Restart CCB
ccb kill -f
ccb

Issue: Stale processes after kill

Solution:

# Force cleanup
ccb kill -f

# If still stuck, find CCB processes
ps aux | grep ccb

# Kill manually
kill -9 <pid>

# Restart
ccb

Best Practices

1. Use environment variables for API keys: Never commit real keys to .ccb/ccb.config
2. Name agents by role: writer, reviewer, tester are clearer than agent1, agent2
3. Use worktrees for isolation: Mark test/experimental agents with (worktree)
4. Update shared memory: Keep .ccb/ccb_memory.md current for team context
5. Explicit delegation first: Use /ask when you know the target; let agents decide only when workflow is clear
6. Start safe: Use ccb -s to preserve manual permission settings during development
7. Clean restarts: Use ccb -n when changing layouts or providers

Configuration Examples

Minimal Two-Agent Setup

.ccb/ccb.config:

writer:codex, reviewer:claude

Complex Multi-Provider Team

.ccb/ccb.config:

cmd; builder:codex, reviewer:claude; qa:gemini(worktree), researcher:gemini

[agents.builder]
key = "$OPENAI_BUILDER_KEY"
model = "gpt-4o"

[agents.reviewer]
key = "$ANTHROPIC_REVIEWER_KEY"
model = "claude-3-5-sonnet-20241022"

[agents.qa]
key = "$GEMINI_QA_KEY"
model = "gemini-2.0-flash-exp"

[agents.researcher]
key = "$GEMINI_RESEARCH_KEY"
model = "gemini-1.5-pro"

Same Provider, Different Endpoints

.ccb/ccb.config:

cmd; prod:codex, staging:codex

[agents.prod]
key = "$OPENAI_PROD_KEY"
url = "https://api.openai.com/v1"
model = "gpt-4o"

[agents.staging]
key = "$OPENAI_STAGING_KEY"
url = "https://staging.example.com/v1"
model = "gpt-4o-mini"

Additional Resources

• Project homepage: https://github.com/SeemSeam/claude_codex_bridge
• Community: Linux.do forum (testing and feedback)
• Contact: bfly123@126.com, WeChat: seemseam-com
• Provider CLIs: Ensure Claude, Codex, Gemini, OpenCode, or Droid CLIs are installed separately

CCB is agent-orchestration infrastructure. It does not bundle agent CLIs—install them separately and configure their API keys via environment variables.