opencode-delegate
Claude Code plugin: delegate implementation to opencode's free models in a git worktree; Claude plans, reviews, merges.
- What
- Claude Code plugin: delegate implementation to opencode's free models in a git worktree; Claude plans, reviews, merges.
- Cost
- Free
- Needs
- Claude Code 2.1 or later (the plugin host); opencode 1.18 or later, or 2.x, installed and authenticated separately (run `opencode auth login` yourself if needed, then check that `opencode providers list` shows OpenCode Zen, where the free models live); git 2.x, jq, curl and bash 4 or later. The plugin carries none of these and carries no credentials: opencode's authentication is your own machine state, and no secret is ever pasted here. In use, your briefs and the code they cover are sent to opencode's free models to be implemented and reviewed, so keep secrets and sensitive code out of the briefs. Nothing is sent anywhere during installation itself, apart from fetching this repository from GitHub.
- Install
- Copy the installer prompt below into your Muse — your agent does the rest.
Curated by Skill Harbor: opencode-delegate is a Claude Code plugin (and plugin marketplace) that runs opencode-driven development. Claude brainstorms with you, writes an implementation plan with bite-sized tasks, and you approve it exactly once. Then, for each task, Claude writes a brief and dispatches an opencode session on a free model (Big Pickle by default) inside a git worktree: the model implements, commits per step and runs the verification command. Claude re-runs verification itself, sends the diff to an opencode reviewer, loops on fixes, and finishes with a whole-branch review before offering to merge, open a PR or leave the branch alone. The skill's own posture: opencode models are capable engineers, and your leverage is the brief. By @Sighton-GH (Bryan Ma), listed here with credit to its creator. Honest caveats: your code is sent to opencode's models to be implemented, so keep secrets and sensitive code out of the briefs; tasks touching auth, secrets, crypto or payments trigger one extra confirmation; between plan approval and the end, only four things stop the run (a destructive operation, a security-sensitive action, a side effect outside the worktree, or a broken plan). Updated October 20, 2026: the built-in default model is now Big Pickle (opencode/big-pickle), because the previous default model left opencode's catalog when opencode 2.0 shipped; the same upstream update (plugin 0.6.0, covering 0.5.0) adds opencode 2.x support, reports provider rate and usage limits as limits with their reset time, denies git stash to opencode sessions, and writes a run summary on every exit path — the workflow, its single plan approval and its four stops are unchanged. Skill Harbor never reviews the code, review it yourself before use.
Version:
Quick answer
opencode-delegate is a free Muse skill, a Claude Code plugin that delegates implementation to opencode's free models in a git worktree, while Claude plans, reviews, and merges. Claude brainstorms with you, writes an implementation plan with bite-sized tasks, and you approve it exactly once, then for each task Claude writes a brief and dispatches an opencode session to implement it.
Install
Copy the install package below, then paste it into Muse🧪 Install ran in our test sandbox · Nov 11, 2026📌 Pinned version: commit 2ed68d3Step 0, before installing anything: read this install prompt, summarize in a few lines what this build does and what installing it involves, then ask me to confirm before you proceed. Prerequisites: Claude Code 2.1 or later (the plugin host); opencode 1.18 or later, or 2.x, installed and authenticated separately (run `opencode auth login` yourself if needed, then check that `opencode providers list` shows OpenCode Zen, where the free models live); git 2.x, jq, curl and bash 4 or later. The plugin carries none of these and carries no credentials: opencode's authentication is your own machine state, and no secret is ever pasted here. In use, your briefs and the code they cover are sent to opencode's free models to be implemented and reviewed, so keep secrets and sensitive code out of the briefs. Nothing is sent anywhere during installation itself, apart from fetching this repository from GitHub. Install "opencode-delegate" for me. It is a Claude Code plugin (this repository is also its plugin marketplace) for opencode-driven development: Claude brainstorms with you, writes the plan and every task brief, re-runs verification, adjudicates reviews and merges, while opencode sessions on free models implement each task, review each task and review the whole branch inside a git worktree, by @Sighton-GH (Bryan Ma), MIT-licensed. Repository: https://raw.githubusercontent.com/Sighton-GH/opencode-delegate/2ed68d3a9d4524c55085d72a821d6f20b431e3a7/plugins/opencode-delegate/skills/opencode-driven-development/SKILL.md (the skill is not at the repository root: the plugin is the plugins/opencode-delegate/ folder at that commit, and the skill inside it is skills/opencode-driven-development/, with SKILL.md, implementer-brief.md, task-reviewer-brief.md, re-review-brief.md, final-reviewer-brief.md and plan-template.md; the plugin also ships bin/oc-task, bin/oc-models and bin/oc-undo, which the skill drives) Pinned version: install the code at commit 2ed68d3a9d4524c55085d72a821d6f20b431e3a7 exactly (Skill Harbor tested this version). Do not substitute a newer version. If this commit is unreachable, stop and tell me. Updated October 20, 2026: Skill Harbor regenerated this install prompt from the current upstream files, at plugin version 0.6.0. The change announced for this update is real, and it is not the only one: the built-in default model changed from Muse Spark 1.3 to Big Pickle (opencode/big-pickle) — in the skill's frontmatter, its posture and model-selection sections, both READMEs and the oc-task script itself — because opencode/muse-spark-1.3-contributor-free left the catalog when opencode 2.0 shipped; free model ids rotate, and oc-models --free always shows the live list. The same update range also brings 0.5.0 and 0.6.0: opencode 2.x is now supported alongside 1.18 or later (oc-task detects the major version and, under 2.x, generates a per-run password for its own local opencode server and uses the new session API); provider rate and usage limits are reported as limits, with a new exit 5 behaviour, a new exit 7 for a busy opencode database, and a run summary and record written on every exit path; git stash is denied to every opencode session; --base, --continue-from and --install were added; and a fresh worktree now gets its own node_modules directory of per-entry links instead of a single link. Two upstream quirks are disclosed rather than smoothed over: the plugin.json description still says "Muse Spark 1.3 by default" even in 0.6.0 (stale metadata — the skill, the READMEs and the code use Big Pickle), and marketplace.json still says version 0.5.0 while plugin.json says 0.6.0. 1. Fetch the whole repository at the pinned commit into a temporary folder, then compare each installed file with the fingerprints from Skill Harbor's review of 2026-10-20: [FINGERPRINTS] - plugins/opencode-delegate/skills/opencode-driven-development/SKILL.md: SHA-256 2831b2b97abea1f52384213f445768082821e314e3f1af8679e75372cffed38c - plugins/opencode-delegate/skills/opencode-driven-development/implementer-brief.md: SHA-256 4d53d9a790899c650f3846dbbdf7ed15fa205ef6b92b98f77fd082944c77edb8 - plugins/opencode-delegate/skills/opencode-driven-development/task-reviewer-brief.md: SHA-256 a02fb87cee4a4456e43f702114da863ba9b0090ca24c1bbddcc027ca82ab083f - plugins/opencode-delegate/skills/opencode-driven-development/re-review-brief.md: SHA-256 6f187741ff53b4ba1d1e85f8b64ea948dcd561f6fb53df76397da93a137a4480 - plugins/opencode-delegate/skills/opencode-driven-development/final-reviewer-brief.md: SHA-256 34c79807d40ed757cdee2b7ca6df37d67445d03aa136ee413c448bb38aa20975 - plugins/opencode-delegate/skills/opencode-driven-development/plan-template.md: SHA-256 738032cec590bf4d583e67f02d7f798c53c0b68cf71f7e77d66068e5a3ab8927 - plugins/opencode-delegate/bin/oc-task: SHA-256 0549bfed6e4af20d9973e233d7ec17337d160d8873a32eeb50ec4f12305eb6c7 - plugins/opencode-delegate/bin/oc-models: SHA-256 9e21587fa14e8c9c1510084bab16c480d510ddeb4a19418906891629f8114e62 - plugins/opencode-delegate/bin/oc-undo: SHA-256 98c0e264c22b56f92ffa41555ce3984e1132e9958357eba8cdad6f0892463b0a - plugins/opencode-delegate/.claude-plugin/plugin.json: SHA-256 1ee406de9ebabe539a1f9616a0310ea71778d87de16bb46d704f0c62e9ba9366 - plugins/opencode-delegate/README.md: SHA-256 f2100109393a186902bdce00a8c18aa03255cd48da0ac5cf4382b357c7b80324 - .claude-plugin/marketplace.json: SHA-256 82d3cab1d5381ba7d5997abd2e3a914fa2f2aa73804c86cca4ae62aaacb932c2 If a file does not match, it changed since the review: do not install it yet, tell me what you found and let me decide. 2. Safety check: read SKILL.md, the five companion files, plugin.json, marketplace.json and the three bin/ scripts and review them for anything suspicious (unexpected network calls, shell commands, credential harvesting); STOP on any red flag and tell me. Tell me in plain words what the plugin does when used, which services it connects to, whether it reads or sends any personal information or keys, and anything that seems unusual. Confirm these points from the files: - This is a real plugin with executable bash scripts, not just a text skill: bin/oc-task, bin/oc-models and bin/oc-undo run shell, git, jq and curl on your machine whenever the skill is used. Installing is only registering the verified copy with Claude Code; nothing runs and nothing is sent during installation itself. - In use, oc-task starts a local opencode server on 127.0.0.1 (a random local port) and talks to it with curl. Under opencode 2.x that server requires basic auth, so oc-task generates a new password for each run on your own machine and passes it to the server and to its own client through environment variables; it is not one of your credentials, it is never typed by you, and it goes nowhere else. The code and briefs themselves are then sent by opencode to the model provider you are authenticated with (OpenCode Zen for the free models) — that sending is the working principle of this plugin, and it happens in use, not at install. - The plugin carries no credentials and asks for none. opencode's own authentication stays in your home directory, created by your own `opencode auth login`; never hand a token, a password or an API key to the agent. Tasks touching authentication, session handling, secrets or credentials, cryptography, or payments and billing get one extra question first — whether you are OK sending that code to an opencode model — once per plan. - Implement sessions run with auto-approved permissions only inside a throwaway git worktree, with one denial: git stash, in any form, because the stash is shared by every worktree of the repository. Review sessions run opencode's read-only plan agent. The base branch is untouched until a merge, which requires your verification command as an attestation and refuses files outside the briefs; oc-undo reverts such a merge in one command. One live run per worktree is enforced, and the skill's only stops after plan approval are a destructive operation, a security-sensitive action, a side effect outside the worktree, or a broken plan. - Default and fallbacks: model precedence is --model, then $OC_DELEGATE_MODEL, then model= in ~/.config/oc-delegate/config, then the built-in opencode/big-pickle; oc-task refuses any model id that is not in opencode's live catalog, and it never picks a paid model on its own. Note the stale plugin.json description disclosed above, and that marketplace.json (0.5.0) lags plugin.json (0.6.0). 3. When I say yes, install it as a Claude Code plugin from the verified local copy, so the pinned commit is what actually gets installed: register the fetched clone as a marketplace with `claude plugin marketplace add "<path to the verified clone>"`, then install with `claude plugin install opencode-delegate@opencode-delegate`. Do not use the one-line GitHub marketplace install for this pinned listing: it fetches the current upstream version, not this commit. Then start a new Claude Code session, because plugins load at startup. 4. Verify with no network calls: plugin.json and marketplace.json parse, the skill's frontmatter is valid, SKILL.md, its five companion files and the three bin/ scripts are in place in the installed plugin, and the plugin shows as installed and enabled for the new session. 5. Report what was installed, where, and what I still need to do myself (e.g. have opencode installed and authenticated with OpenCode Zen listed in `opencode providers list`, where the plugin will run; in the new session, ask Claude to run `oc-models --free` as the live check — a list of provider/model ids means the plugin is live, "command not found" means it is not enabled for that session; optionally install the superpowers plugin as a companion for brainstorming and plan writing, the skill has built-in fallbacks without it; to stop Claude from starting the skill on its own, set `disable-model-invocation: true` in the skill's SKILL.md). GitHub is optional: if I have a GitHub account or the gh CLI, you may use it; otherwise public access is fine. Never require it unless it's in the prerequisites above. Rules: don't touch anything outside the temp folder and the install target. If anything looks off, stop and ask me.
Saved to your recent installs. Find it anytime on /connect.
🔒 Security, provenance and what this touches
- It can run commands on your Muse’s computer.
What the scan found — excerpts from the install prompt
- runs commands — Prerequisites (the plugin does not carry any of these): Claude Code 2.1+ (`claude --version`); opencode 1.18+ · Prérequis (le plugin n'en fournit aucun) : Claude Code 2.1+ (`claude --version`) ; opencode 1.18+ INSTALLÉ ET
The install prompt below already includes the vetting steps: your agent follows the community checklist before installing anything with executable code. Want more?
Verification certificate
Dated facts about what was checked for this listing, and how. This is not a safety certification.
- Pinned version
- 2ed68d3 · 2026-10-20
- Sandbox install test
- Install ran in our test sandbox · Nov 11, 2026
- Creator identity
- Not verified
- Fingerprint watch
- Last checked · Nov 19, 2026 : no change detected in the files linked by this install prompt
- Install prompt analysis
- 2026-10-13 : runs commands
Skill Harbor never reviews the code of a listed skill. These are dated facts about the install prompt and our install test, not a safety certification. Review a skill yourself before you install it. How we verify
Questions
How do I install a skill?
Every product page includes a copy-paste install prompt. Paste it into your Muse and it sets the skill up for you — no manual configuration.
Where does my money go?
Straight to the seller. Skill Harbor never processes payments: checkout happens on the seller’s own page, usually Stripe.
What does the ✓ next to a creator’s name mean?
It means we confirmed the identity of the person behind the listing. It says nothing about the code itself — always check a skill before installing it.