Skip to main content

opencode

Timon can run agents on opencode, the open-source coding agent, including its free Muse Spark 1.3 model, which needs no account. It drives the opencode CLI over the Agent Client Protocol; opencode keeps its own sign-in and provider keys, and Timon never sees or stores them.

Set it up​

  1. Install the CLI:

    npm install -g opencode-ai

  2. Optional: to use your own providers instead of the free model, sign in in a terminal, under the same OS account that runs Timon:

    opencode auth login

  3. Restart Timon. opencode then appears among the LLM backends.

Timon looks for opencode on the service's PATH, in the npm and Homebrew folders, and in ~/.opencode/bin, where opencode's own install script puts it; on Windows it also looks in the Scoop (%USERPROFILE%\scoop\shims) and Chocolatey (%ProgramData%\chocolatey\bin) folders.

If opencode does not appear, run timon doctor: it tells you whether the opencode binary was found. A CLI installed anywhere else may be missing from the service's PATH; set host.binaries.opencode to its absolute path in that case. If a mission stops with a sign-in message, run the command it names in a terminal on the machine that runs Timon, then retry.

Models​

The default model is Muse Spark 1.3 (free), which opencode offers without an account. The list shown in the model picker comes from opencode itself, so it follows the providers you signed in to, and it is refreshed at most every 12 hours. Until the first answer, only the free model is listed.

Opening that list starts opencode briefly, which leaves one empty session in opencode's own history. A model change applies from the next mission or conversation, not in the middle of one.

What agents can do​

opencode decides its own tool permissions. Timon shows no Allow / Deny card for it and cannot restrict what it does inside your project. opencode refuses, for example, a write outside the project folder on its own, without asking. The agent editor flags this with a "manages its own permissions" badge. Choose opencode for an agent you are comfortable letting work with opencode's own rules.

Replies stream into the chat, tool calls appear with their results, and the model's reasoning shows as a collapsed Reasoning block. Stop in a chat ends the current turn and keeps the conversation; the next message continues in the same opencode session, including after Timon restarts. Timon's own tools (Kanban, questions, missions, memory) and the stdio MCP servers configured for the agent reach opencode.

Limitations​

  • Permissions are opencode's. No Allow / Deny cards, and the agent's permission mode does not narrow what opencode may do.
  • HTTP MCP servers are not passed on. MCP servers you added over HTTP do not reach opencode agents.
  • No live model switch. A model change applies from the next session.
  • No quota gauge. Timon does not display your remaining usage with opencode's providers.
  • Not yet tried on every setup. Windows and macOS installs in particular have had little testing; report what you see.