Enhancement Official

OpenCode (Provider)

Run NanoClaw agents on OpenCode instead of the Claude Agent SDK. OpenRouter, Anthropic, DeepSeek, Google, and any OpenAI-compatible provider.

What it does

Replace Claude Agent SDK with OpenCode as the agent provider
Works with OpenRouter, Anthropic, DeepSeek, Google, OpenCode Zen, and others
Per-group selection via container.json — mix providers freely
OneCLI credential proxy supported — keys never live in .env
MCP tools shared across providers (same merged config)

What you'll need

NanoClaw installed and running
OneCLI Agent Vault for credential injection (recommended)
An API key from at least one supported provider

Install

/add-opencode

How it works

NanoClaw runs each agent group in a long-lived poll loop inside its container. The backend is selected with AGENT_PROVIDER — claude (default), opencode, codex, or mock.

The /add-opencode skill copies the OpenCode provider from the providers branch and wires it in. OpenCode keeps a local opencode serve process and SSE subscription. Each turn opens a new stream; teardown is stream.return() plus a SIGKILL on the server process to avoid zombie / MCP hangs.

OpenCode is multi-provider. You set a provider id (anthropic, openrouter, deepseek, opencode/Zen, etc.) and a model id, and OpenCode handles the dispatch.

Setup

The skill walks you through:

Fetch the providers branch.
Copy opencode.ts and mcp-to-opencode.ts into the host and container provider trees.
Add the self-registration import to both barrels.
bun add @opencode-ai/sdk@1.4.17 in container/agent-runner/ (must match the CLI version pinned in step 5).
Pin OPENCODE_VERSION in container/Dockerfile and add opencode-ai@${OPENCODE_VERSION} to the global pnpm install.
Rebuild host and container.

If you have existing agent groups, the skill also propagates the new files into each group’s data/v2-sessions/<id>/agent-runner-src/providers/ overlay — the runtime overlay overrides the image, so it has to be updated separately.

Provider configurations

A few common shapes:

OpenRouter:

OPENCODE_PROVIDER=openrouter
OPENCODE_MODEL=openrouter/anthropic/claude-sonnet-4
ANTHROPIC_BASE_URL=https://openrouter.ai/api/v1

DeepSeek:

OPENCODE_PROVIDER=deepseek
OPENCODE_MODEL=deepseek/deepseek-chat
ANTHROPIC_BASE_URL=https://api.deepseek.com/v1

Anthropic (no ANTHROPIC_BASE_URL needed):

OPENCODE_PROVIDER=anthropic
OPENCODE_MODEL=anthropic/claude-sonnet-4-20250514

OpenCode Zen uses x-api-key headers, not Bearer — register the OneCLI secret with --header-name "x-api-key" and --value-format "{value}".

For each provider, register the API key with OneCLI using a matching --host-pattern so the credential proxy injects it for the right outbound calls.

Per-group config

Set the provider in two places that must agree:

groups/<folder>/container.json → "provider": "opencode"
agent_groups.agent_provider → 'opencode' (or per-session override via sessions.agent_provider)

The container reads container.json directly. The DB columns drive host-side env passthrough (per-session XDG mount, OPENCODE_* env).

Tips

Pin both the SDK and the CLI to 1.4.x. The 1.14.x SDK has a completely different API and breaks the provider code. latest will silently upgrade to a broken version.
Build cache gotcha: if you see “Unknown provider: opencode” after a rebuild, the buildkit cached an older COPY. Run docker builder prune -f && ./container/build.sh.
NO_PROXY matters when the OpenCode client talks to 127.0.0.1 inside the container while OneCLI’s HTTPS_PROXY is set.
Naming overlap: AGENT_PROVIDER=opencode (NanoClaw setting) means “use the OpenCode runner.” OPENCODE_PROVIDER=opencode (OpenCode’s setting) means “use Zen as the upstream provider.” Yes, that’s confusing.
Codex vs. OpenCode: Codex is OpenAI-only with a simpler config. OpenCode is multi-provider but has more knobs. Pick based on which model selection style fits.