Opencode Controller: Let Clawdbot Drive Opencode — Without Writing a Single Line of Code

12,700+ downloads and 55 stars — the Opencode Controller Skill by @Karatla bridges two AI coding tools: Clawdbot (your conversational AI assistant) and Opencode (a dedicated AI coding agent). The premise is unusual and powerful: Clawdbot doesn't write the code. Opencode does. Clawdbot orchestrates the whole process.

What is Opencode?

Opencode (opencode.ai, GitHub: opencode-ai/opencode) is an open-source AI coding agent built for the terminal, created by Anomaly Innovations. It competes directly with Claude Code and GitHub Copilot and has become one of the most-watched open-source coding agents in early 2026.

Key facts about Opencode:

• 75+ LLM providers — Claude, OpenAI, Gemini, local models via LM Studio, and more
• Provider-agnostic — no vendor lock-in; switch providers per session
• Two built-in modes — Build (full-access implementation) and Plan (read-only analysis)
• Privacy-first — does not store your code or context; you control all session sharing
• Multi-surface — terminal TUI, desktop app, and extensions for VS Code, Cursor, JetBrains, Zed, and Neovim
• LSP integration — Rust, Swift, TypeScript, Python, Terraform, and more

The Opencode Controller Skill gives Clawdbot the workflow knowledge to orchestrate Opencode via its slash commands.

The Core Rule

Clawdbot does not write code. All planning and coding happens inside Opencode.

This is the fundamental design principle of the skill. Instead of having two AI systems independently generate code (with inconsistency and context loss), Clawdbot acts as the intelligent operator — asking the right questions, managing the workflow, and routing tasks to Opencode at the right moment.

The Standard Workflow

1. Confirm AI provider and auth method with the user
2. Start Opencode
3. Reuse existing project session (or create new if needed)
4. Select AI model via /models
5. Enter Plan mode
6. Generate and validate the plan with the user
7. Switch to Build mode
8. Implement
9. Return to Plan mode for any unexpected complexity
10. Repeat until finished

Key Commands

Session Management

# Open session selector in Opencode
/sessions
 
# Rule: Same project ALWAYS uses the same session.
# Never create a new session without explicit user approval.
# Sessions preserve project context and decisions.

Mode Switching

# Open agent selector
/agents
 
# Available modes:
# - Plan: thinking, architecture, questions, no code
# - Build: code implementation only
 
# Switch modes with Tab key
# Always start in Plan mode

Model Selection

# Open model selector
/models
 
# For auth-based providers (OpenAI, etc.):
# 1. Opencode generates a login URL
# 2. Copy and send to user
# 3. Wait for confirmation BEFORE continuing
# Never assume authentication is complete

Plan vs. Build: The Critical Distinction

The golden rule: Never switch to Build mode until the plan is explicitly approved. If Opencode asks a question during Build mode, immediately switch back to Plan mode to answer it.

Failure Handling

Usage Example

With the skill installed, a typical Clawdbot session looks like:

You: "I need to add authentication to my Express app"

Clawdbot: Which AI provider should I use in Opencode?

You: "OpenAI"

Clawdbot: How should Opencode authenticate — auth, key, or local?

[Clawdbot starts Opencode, opens session selector, finds existing project,
selects model, enters Plan mode, asks Opencode to analyze the task]

Clawdbot: Opencode proposes: [plan details]. Shall I proceed to Build mode?

You: "Yes, looks good"

[Clawdbot switches to Build mode, Opencode implements]

Why Two AI Systems?

You might wonder: why run Clawdbot and Opencode? The answer is specialization:

• Clawdbot excels at natural language understanding, user interaction, and workflow orchestration
• Opencode excels at code generation with project-level context, multi-file awareness, and multiple AI provider support

The Opencode Controller Skill makes these complementary strengths work together — Clawdbot handles the conversation and coordination, Opencode handles the actual implementation.

Considerations

• Opencode must be installed separately — the skill controls Opencode but doesn't install it; install Opencode first
• Session discipline matters — always reusing the correct session is critical; starting new sessions loses project context
• Auth patience required — when providers require web authentication, the workflow pauses for user confirmation; this is intentional and important
• Strict mode enforcement — the skill enforces the Plan→Build sequence rigidly; this is a feature, not a limitation — it prevents the common failure of jumping to implementation before fully understanding requirements

The Bigger Picture

The Opencode Controller Skill represents a new pattern in AI tooling: AI agents orchestrating other AI agents. Instead of one monolithic AI system trying to do everything, specialized agents collaborate — each doing what it does best. With 12,700+ downloads and 55 stars, developers have found real value in this separation of concerns. It's a preview of how complex AI workflows will be structured: not single systems, but coordinated ensembles.

View the skill on ClawHub: opencode-controller