Hivemind

Claude Code plans and reviews. Free opencode workers do the grunt work.

Your expensive model's context is the scarce resource — not its intelligence. Hivemind moves mechanical work (multi-lens review, research sweeps, bulk migrations, partitioned test runs) onto disposable opencode workers running on free models, and returns exactly one compact JSON line per worker. Raw agent streams never touch the orchestrator's context.

The orchestrator stays the only planner, the only reviewer, and the only merger. Workers are cattle, not colleagues.

/hive migrate every fetch() call in src/ to the shared http client
   |
   +-- worktree-1  coder   -> { ok: true, tokens: 8214, cost_usd: 0, duration_ms: 41203 }
   +-- worktree-2  coder   -> { ok: true, tokens: 6980, cost_usd: 0, duration_ms: 38771 }
   +-- worktree-3  tester  -> { ok: true, result: "PASS 214/214", ... }
   |
   +-- you review every diff, you merge

Install

As a Claude Code plugin:

/plugin marketplace add Hanishchow/hivemind
/plugin install hivemind@hivemind

Or by hand:

git clone https://github.com/Hanishchow/hivemind
cp -r hivemind/skills/hivemind ~/.claude/skills/
cp hivemind/commands/*.md ~/.claude/commands/
cp hivemind/skills/hivemind/assets/agents/*.md ~/.config/opencode/agent/
export HIVEMIND_HOME="$HOME/.claude/skills/hivemind"

Prerequisites

Hivemind is an orchestration layer, not a model. It needs:

Requirement	Notes
Node.js >= 18	The scripts use fetch and node:timers/promises.
opencode CLI on PATH	Install and authenticate it yourself.
Default model opencode/mimo-v2.5-free	A free tier offered by opencode, not by Anthropic. Availability, rate limits, and pricing are theirs to change — override with --model.
Windows: OPENCODE_GIT_BASH_PATH	Point at C:\Program Files\Git\bin\bash.exe, set persistently.

Worker traffic goes to opencode's endpoints. Don't delegate secrets or private code you wouldn't send there.

Commands

Command	Shape
/hive <task>	Auto-router — classifies the task, then picks one of the below. Start here.
/oc <task>	One worker, one task.
/swarm <task>	2-5 independent subtasks in parallel, each in its own git worktree.
/review-panel <diff>	Four reviewers (correctness / security / performance / style), consensus-filtered.
/research-sweep <question>	3-5 scouts on different angles, synthesized, every claim cited.
/migration <task>	Batched per-worktree migration workers with a sequenced merge.
/test-fleet <target>	A test suite partitioned across parallel tester workers.

Three worker personas ship with it: scout (read-only, cannot write or run shell), coder (writes, confined to one worktree), tester (runs tests, never edits source).

The contract

Every worker is invoked through one hardened entry point, and returns one line:

node "$HIVEMIND_HOME/scripts/oc-worker.mjs" --agent scout --run r1 --label auth "Where is the session token validated?"

{"ok":true,"result":"...","tokens":{"total":8214},"cost_usd":0,"duration_ms":41203,"label":"auth","agent":"scout"}

Failures return the same shape: { ok: false, stage: "args"|"exec"|"api"|"parse"|"empty", error } with stderr capped at 300 chars. Nothing else reaches the orchestrator — no NDJSON, no stack traces, no partial streams. --run / --label append lifecycle events to .runs/<id>.jsonl, so oc-status.mjs can recover fleet progress even after the orchestrator loses its context.

Fallback ladder

1. Worker returns ok:false -> re-invoke once against the same directory.
2. Still failing -> the orchestrator does that subtask itself, marked [orchestrator-sourced].
3. opencode down entirely -> announce it, abandon the workers, do the task directly.

A swarm should never fail a task the orchestrator could have done alone.

Benchmarking

scripts/bench/ ships an A/B/C harness — A: Claude solo, B: opencode solo, C: Claude orchestrating two workers — appending (ts, config, tokens, cost, duration) records to bench-results.jsonl. Grade the artifacts blind with scripts/bench/grader-prompt.md, which hides everything but the task spec and the output. Run it on your own workload; results depend far more on your task mix than on anything published here.

Known limits

• Free-tier rate limits will 429 under heavy swarms. Space out retries.
• Worker quality varies. Review every diff — that's why the orchestrator is the only merger.
• Config C still spends real orchestrator tokens (~1-2k/task) on coordination.
• opencode-go/* models require workspace billing. Avoid them if you want the $0 path.

Status

Also submitted upstream, both open at the time of writing: anthropics/skills#1628 and alirezarezvani/claude-skills#979.

License

Apache 2.0 — see LICENSE.