Opencode

opencode plugin for Claude Code

Use opencode from inside Claude Code for code reviews or to delegate tasks to opencode.

This plugin is for Claude Code users who want an easy way to start using opencode from the workflow they already have.

Inspired by and structurally modeled on openai/codex-plugin-cc — same slash-command shape, same companion-script architecture, retargeted at opencode's REST server instead of Codex's JSON-RPC app server.

What You Get

• /opencode:review for a normal read-only opencode review
• /opencode:adversarial-review for a steerable challenge review
• /opencode:rescue to hand a free-form task to opencode through a subagent
• /opencode:status, /opencode:result, /opencode:cancel to manage tracked background jobs
• /opencode:broker and /opencode:sessions to control the long-lived opencode serve process and list its sessions
• /opencode:setup to verify the opencode CLI is installed

Requirements

• An opencode-compatible model. opencode supports any provider on models.dev — Anthropic, OpenAI, OpenRouter, Moonshot, GitHub Copilot, etc. The plugin doesn't pick a model for you; whatever opencode is configured to use is what you get.
• Node.js 20 or later (for the companion script that ships with the plugin).

Install

Add the marketplace in Claude Code:

/plugin marketplace add JoelFrancisco/opencode-plugin-cc

Install the plugin:

/plugin install opencode@opencode-plugin-cc

Reload plugins:

/reload-plugins

Then run:

/opencode:setup

/opencode:setup will tell you whether opencode is installed. If it isn't, install it with:

curl -fsSL https://opencode.ai/install | bash

If opencode is installed but not authenticated yet, run:

!opencode auth login

After install, you should see:

• the slash commands listed below
• the opencode:opencode-rescue subagent in /agents

One simple first run:

/opencode:review --background
/opencode:status
/opencode:result

Usage

/opencode:review

Runs a normal opencode review on your current work.

[!NOTE] Code review for multi-file changes might take a while. It's generally recommended to run it in the background.

Use it when you want:

• a review of your current uncommitted changes
• a review of your branch compared to a base branch like main

Supports --base <ref> for branch review, --model <provider/model> to override opencode's configured model, plus --wait and --background to choose execution mode. It is not steerable and does not take custom focus text — use /opencode:adversarial-review for that.

Examples:

/opencode:review
/opencode:review --base main
/opencode:review --background
/opencode:review --model anthropic/claude-sonnet-4-6

This command is read-only. When run in the background you can use /opencode:status to check progress and /opencode:cancel to abort.

/opencode:adversarial-review

Runs a steerable review that questions the chosen implementation and design.

It can be used to pressure-test assumptions, tradeoffs, failure modes, and whether a different approach would have been safer or simpler.

Same scope flags as /opencode:review (--base <ref>, --wait, --background, --model). Unlike /opencode:review, it can take extra focus text after the flags — that becomes the Focus areas: line in the prompt.

Examples:

/opencode:adversarial-review
/opencode:adversarial-review --base main challenge whether this caching design is safe
/opencode:adversarial-review --background look for race conditions and question the chosen approach

This command is read-only. It does not fix code.

/opencode:rescue

Hands a task to opencode through the opencode:opencode-rescue subagent.

Use it when you want opencode to:

• investigate a bug
• try a fix
• continue a previous opencode task in this workspace
• take a faster pass with a smaller model

[!NOTE] These tasks can take a long time depending on the model and the work — it's generally recommended to run them in the background.

Supports --background, --wait, --resume, --fresh, and --model <provider/model>. If you omit --resume and --fresh, the plugin offers to continue the latest rescue thread for this workspace when one exists.

Examples:

/opencode:rescue investigate why the tests started failing
/opencode:rescue fix the failing test with the smallest safe patch
/opencode:rescue --resume apply the top fix from the last run
/opencode:rescue --background investigate the regression
/opencode:rescue --model openrouter/moonshotai/kimi-k2-0905 sketch a faster algorithm

You can also just ask for a task to be delegated:

Ask opencode to redesign the database connection to be more resilient.

Notes:

• if you do not pass --model, opencode uses its configured default
• --resume and --fresh are mutually exclusive
• the plugin tracks rescue sessions by their title ("rescue") so resume picks up the latest one in the current workspace

/opencode:status

Shows running and recent opencode jobs for the current workspace.

Examples:

/opencode:status

Use it to:

• check progress on background work
• see the latest completed job
• confirm whether a task is still running

/opencode:result

Shows the final stored opencode output for a finished job.

Examples:

/opencode:result
/opencode:result --job <id>

Defaults to the most recent job in the current workspace.

/opencode:cancel

Cancels an active background opencode job (SIGTERMs the tracked PID and writes status: cancelled).

Examples:

/opencode:cancel
/opencode:cancel --job <id>

/opencode:broker

Manages the long-lived opencode serve process for the current workspace. The broker is started lazily on the first review, kept alive across subsequent calls (no per-call npx/startup tax), and stopped automatically on Claude Code's SessionEnd hook.

Examples:

/opencode:broker start
/opencode:broker status
/opencode:broker stop

Lockfile lives at ${OPENCODE_PLUGIN_STATE_DIR:-~/.local/state/opencode-plugin-cc}/server/.

/opencode:sessions

Lists opencode sessions in the current workspace via the broker.

/opencode:sessions

If no broker is running, you'll be told to start one with /opencode:broker start.

/opencode:setup

Checks whether opencode is installed and on PATH. If it isn't, prints the install hint (curl -fsSL https://opencode.ai/install | bash).

/opencode:setup

Typical Flows

Review Before Shipping

/opencode:review

Hand A Problem To opencode

/opencode:rescue investigate why the build is failing in CI

Start Something Long-Running

/opencode:adversarial-review --background
/opencode:rescue --background investigate the flaky test

Then check in with:

/opencode:status
/opencode:result

opencode Integration

The plugin wraps the opencode HTTP server. It uses the global opencode binary on your PATH and inherits its configuration and authentication.

Common Configurations

If you want to change the default model or provider that gets used by the plugin, set them in opencode's own config:

• user-level: ~/.config/opencode/opencode.json
• project-level: ./opencode.json

The model field there picks the default. The plugin only forwards a model when you pass --model <provider/model> explicitly — otherwise it inherits.

See the opencode config docs for the full set of options.

Resuming Inside opencode

Tracked jobs record the opencode session id (JobState.sessionId). You can reopen any of them inside opencode directly:

opencode run --session <session-id>

Or use the plugin's own /opencode:rescue --resume to continue the latest rescue thread without leaving Claude Code.

--effort

Soft hint passed to /opencode:review, /opencode:adversarial-review, and /opencode:rescue. Accepts the same vocabulary as codex-plugin-cc: none | minimal | low | medium | high | xhigh. Default is unset, which means medium (no extra prompt instruction).

opencode does not have a uniform reasoning-effort API the way codex's --effort does — different providers expose different knobs (Anthropic thinking budget, OpenAI reasoning effort, none for Moonshot). Rather than trying to translate per-provider, the plugin appends a one-line instruction to the prompt (e.g. "Think carefully through edge cases…" for high, "Answer in 1-2 sentences max…" for minimal). Different providers will respond to that instruction differently — it's a heuristic, not a guaranteed model parameter.

If you want hard control over reasoning depth on a specific provider, set it in opencode's config rather than relying on --effort.

Out-of-Scope vs codex-plugin-cc

A few codex-plugin-cc features are intentionally not mirrored here:

• Stop-time review gate — codex prompts for a review on session stop; this plugin doesn't (the loop risk and usage drain aren't worth it for most workflows).
• --write — opencode's tools are write-capable by default, no analog needed.
• Model aliases — codex's spark → gpt-5.3-codex-spark-style mappings. Pass full provider/model strings.

FAQ

Do I need a separate opencode account?

No. The plugin uses whatever provider you've already authenticated with via opencode (opencode auth login).

Does the plugin use a separate opencode runtime?

No. It delegates through your local opencode CLI on the same machine. Same install, same authentication state, same configuration.

Does the plugin pick a model?

No — model selection lives entirely on the user's command line (--model provider/model) or in opencode's own config. The plugin never defaults a model and never aliases one.

Development

pnpm install
pnpm build         # tsc → dist/
pnpm typecheck
pnpm lint          # oxlint
pnpm fmt:check     # oxfmt
pnpm test          # vitest (Layers 0 + A)
pnpm test:e2e      # adds Layer B (Claude Code dispatch)
pnpm test:live     # adds Layer C (real opencode + real model)

The compiled dist/ is committed so the plugin works without a build step at install time. Don't forget to rebuild before pushing.

See AGENTS.md for architecture notes and the build-order breakdown.

License

MIT.