OpenCode — run your chats on a different engine

MeshWeaver chat runs each round of a conversation through an execution harness: the engine that receives the prompt, calls tools and writes the answer. This package adds OpenCode as one of those harnesses. Pick it in a thread and the round runs on the opencode command-line agent, driven over the Agent Client Protocol (ACP), while the thread, the mesh tools, the nodes and the context stay exactly what they were.

It is for people who want to try another agent engine on their own work without leaving the portal: OpenCode's free models are available without any key, and its paid models are one login away.

Key concepts

• A harness, not an agent. A harness changes what executes the round, not which agent you talk to. Your agents, skills and instructions remain the same; only the engine underneath changes. The OpenCode harness does not offer its own agent selection.
• One thread, one OpenCode session. Every round of a thread resumes the same OpenCode session, so the agent keeps the thread's context and earlier tool results itself rather than having the whole history replayed.
• Visible tool calls. Every tool the agent runs shows up in the thread as a tool call, the same way it does with the standard MeshWeaver agent path.
• Shared ACP driver. The code that speaks the Agent Client Protocol is shared by every ACP agent and ships with the ACP Agents package (Acp). This package is what puts OpenCode into your picker; it declares ACP Agents and the AI engine as dependencies, so the installation resolves them for you.

A typical workflow

1. Get the package. Once you hold it, the harness appears in the harness picker of your threads — the picker lists the harnesses of every package you can read.
2. Open a thread and pick OpenCode as the harness.
3. Chat right away on OpenCode's free models — or run /login in the thread and paste an OpenCode Zen API key to use the paid ones.
4. Work as usual. Ask questions about your nodes, let the agent search, read and edit through the mesh tools you already have, and follow each tool call in the thread.

Threads keep working whichever harness answered a given round: they are ordinary mesh nodes, so the conversation stays readable, searchable and shareable afterwards.

What stays private

• No shared key from the mesh. The harness appears only for people who hold the package, and nothing is billed to the mesh.
• Private logins need a per-user root. Your OpenCode login is kept apart from other users' only when the installation sets Acp:ConfigDirRoot (or its fallback ClaudeCode:ConfigDirRoot): each round then runs with its own home, {root}/{your user id}/…. Without that setting the harness uses the host account's own home, so every user of that portal shares whatever OpenCode login is cached there — fine on a single-user machine, not on a shared portal.
• Stopping is explicit. Uninstalling removes only a legacy copy in your home, if one exists, and keeps your access, so the harness stays in your picker. To stop using OpenCode, switch the thread to another harness explicitly; the thread and its history stay as they are.

What an installation needs

The OpenCode CLI itself must be installed beside the portal that runs your threads — the harness starts it and talks to it over ACP. On an installation without the CLI the harness cannot run a round. A portal that several people use also needs Acp:ConfigDirRoot (or ClaudeCode:ConfigDirRoot) pointing at a per-user volume, so that each person's login stays their own.

What you get when you subscribe

The package is in the Pro tier. Holding it adds the OpenCode entry to your harness picker and everything described above; it does not change anything for users who do not hold it.