Feature request: ACP client mode — drive external ACP agents (Codex, Claude Code) from OpenCode

Summary

OpenCode currently implements the agent side of the Agent Client Protocol: opencode acp lets any ACP-compatible editor (Zed, JetBrains AI Assistant, VS Code ACP client) drive OpenCode as an agent.

This request is for the complementary client side: let OpenCode connect to external ACP agents — e.g. Codex via @agentclientprotocol/codex-acp and Claude Code via @agentclientprotocol/claude-agent-acp — and delegate work to them, the same way an editor delegates to OpenCode today.

Motivation

The ACP agent ecosystem is consolidating under the agentclientprotocol org, with officially maintained adapters:

• Codex: @agentclientprotocol/codex-acp wraps the Codex App Server as an ACP agent (auth, model/sandbox config, streamed events, subagent sessions).
• Claude Code: @agentclientprotocol/claude-agent-acp exposes the Claude Agent SDK as an ACP agent (edit review, permission requests, todo lists, subagent transcripts).

Other harnesses are already building ACP-client orchestration on top of these adapters (e.g. QwenPaw's dual server/client ACP support, Hermes' ACP client proposals, Warp's ACP roadmap).

For OpenCode users this unlocks agent-level integrations that MCP alone doesn't cover well today:

• Cross-model-ecosystem second opinions: ask Codex or Claude Code to review a plan/diff from inside an OpenCode session.
• Delegating specialized subtasks to an external agent while OpenCode orchestrates.
• A uniform delegation surface: ACP already models sessions, streaming, permissions, and edit review — semantics OpenCode's own agent-side implementation already speaks, so the client side would reuse the same primitives from the opposite end.

For context, the MCP-based alternatives for this use case are narrowing: OpenAI removed the codex mcp-server (host-as-MCP) command (the migration target is the app-server protocol — which is what codex-acp wraps), and claude mcp serve exposes tool-level rather than agent-level semantics. ACP is the standardized agent-level option, and its adapters are officially maintained.

What it could look like

Rough sketch, purely for discussion:

{
  "acp": {
    "agents": {
      "codex": { "command": ["npx", "-y", "@agentclientprotocol/codex-acp"] },
      "claude": { "command": ["claude-agent-acp"] }
    }
  }
}

External agents could then appear as invocable agents (delegable from the orchestrator, usable as council members, etc.), with OpenCode acting as the ACP client: creating sessions, streaming events, handling permission requests, and rendering edits for review.

Concrete use case

I run OpenCode as my primary harness and keep Codex installed alongside (authenticated via the desktop app). I'd like to say "ask Codex to review this refactor" in an OpenCode session, have the orchestrator delegate to codex-acp, stream its reasoning, surface its edits for approval, and fold the conclusion back into the session — without leaving OpenCode and without a hand-rolled wrapper around codex exec.

Thanks for considering!