opencode-claude-code-tool

OpenCode plugin that exposes a claude_code tool, letting any model delegate agentic coding work to the Claude Code CLI.

Why

Run a cheap orchestrator model (e.g. GLM, Qwen, GPT) for planning, exploration, and decisions, while a Claude model (Sonnet / Opus / Haiku) does the heavy execution — reading, editing, running commands, iterating — in an isolated headless subprocess with its own tool loop.

This is different from just binding a sub-agent to the Claude model: here Claude Code is a full agent (own context, tools, iteration budget), not a single LLM call. The orchestrator stays tiny; Claude does the real work and returns a result.

orchestrator (any model) ──tool call──▶ claude_code({ prompt, ... })
                                               │
                                  claude -p --output-format json \
                                    --model sonnet --permission-mode acceptEdits \
                                    --add-dir <worktree>
                                               │
                          ◀── result text + cost/duration metadata ──

Install

Requires the claude CLI in PATH (npm i -g @anthropic-ai/claude-code or the official installer) and an authenticated Claude account / ANTHROPIC_API_KEY.

As an npm plugin (recommended for sharing)

Add to your opencode.json:

{
  "plugin": ["opencode-claude-code-tool"]
}

OpenCode installs it automatically via Bun on startup.

From a local file

Copy plugin.ts into:

• ~/.config/opencode/plugins/claude-code.ts (global), or
• .opencode/plugins/claude-code.ts (project-level)

Local plugin files are auto-loaded on startup — no config change needed.

Tool: claude_code

Arg	Type	Default	Description
prompt	string (req)	—	The task. Be specific; Claude has its own full tool loop (Edit, Bash, Read, Write) and iterates autonomously.
model	sonnet|opus|haiku|fable	sonnet	Claude model alias.
cwd	string	session worktree	Working dir for the subprocess (scoped via --add-dir).
systemPrompt	string	—	Replace Claude's default system prompt.
appendSystemPrompt	string	—	Append to the default system prompt (preferred).
allowedTools	string[]	all	Restrict tools, e.g. ["Edit","Read","Bash","Write"].
disallowedTools	string[]	—	Deny specific tools, e.g. ["Bash(git push*)"].
maxTurns	number	—	Cap agentic turns; prevents runaway loops.
mcpConfig	string	—	Path/JSON for --mcp-config (extra MCP servers for Claude).
permissionMode	default|acceptEdits|plan|bypassPermissions|auto|dontAsk	bypassPermissions	Claude --permission-mode. Default gives full execution parity with opencode subagents (edits AND bash run without prompting, matching bash: * allow). plan = read-only; acceptEdits = edits only (bash denied in headless).
timeoutMs	number	600000 (10 min)	Hard timeout; raises a clear error if exceeded.

Returns: Claude's final result text. Cost (USD), duration, and turn count are attached as tool metadata.

Enabling for sub-agents

Plugin tools are globally available. To let a specific sub-agent use it, add it to that agent's tools map in opencode.json.

Name sub-agents by their driving model (the one that runs the reasoning loop), not by the executor — the executor (Claude Code) is a tool the agent uses, not its identity. So a GLM-driven apply phase that delegates coding to Claude is sdd-apply-glm, not sdd-apply-noclaude (that suffix means "not bound to Claude as a model" and becomes misleading once it calls Claude):

"sdd-apply-glm": {
  "description": "Implement tasks (GLM-driven, delegates coding to Claude Code)",
  "mode": "subagent",
  // no "model" → runs on the default runtime model (e.g. GLM)
  "prompt": "...",
  "tools": {
    "bash": true,
    "edit": true,
    "read": true,
    "write": true,
    "claude_code": true      // ← delegate the actual coding to Claude Code
  }
}

Now a GLM/Qwen orchestrator can run the apply phase and call claude_code to hand the actual coding to Claude.

Configuration (env)

Var	Default	Description
CLAUDE_BIN	claude (from PATH)	Path to the Claude CLI.
CLAUDE_CODE_MODEL	sonnet	Default model when the tool omits model.
CLAUDE_CODE_PERMISSION_MODE	bypassPermissions	Default --permission-mode.
CLAUDE_CODE_TIMEOUT_MS	600000	Default timeout in ms.

Safety & permissions

The default bypassPermissions mode gives the Claude subprocess the same execution power as an opencode subagent — opencode subagents run with bash: * allow, so they can edit files AND run commands (tests, builds, etc.). bypassPermissions mirrors that: the subprocess can edit and run bash without being prompted (headless -p mode has no TTY to prompt anyway, so any mode short of bypass would deny bash rather than ask).

The subprocess is still scoped to cwd (the session worktree) via --add-dir.

Tighten when needed:

• permissionMode: "plan" — read-only analysis, no edits/bash.
• permissionMode: "acceptEdits" — edits only; bash denied (use when you don't trust the subprocess to run commands).
• allowedTools: ["Edit","Read"] — explicit allowlist.
• disallowedTools: ["Bash(git push*)"] — deny specific dangerous ops while keeping everything else (mirrors opencode's git push: ask guard).

Change the global default via CLAUDE_CODE_PERMISSION_MODE.

The subprocess respects the caller's abort signal and the timeoutMs cap.

License

MIT