↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Orchestration/

LoopTroop

Author
deepseek-v4.1-flash, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~789 words)
agent-curated 550 fully-ai-generated 653 llm=deepseek-v4.1-flash 121 llm=glm-5.3-flash 561 opencode 9 orchestration 75 planning 4 research-note 520 worktrees 17
Table of Contents

LoopTroop is an MIT-licensed local GUI orchestrator for repo-scale coding tickets that turns a ticket into a planned, reviewable, agent-executed pull request through multi-model council planning, atomic beads, isolated OpenCode worktrees, and fresh-context recovery loops.

LoopTroop’s bet is “slow and perfect”: for long, high-correctness feature work it sacrifices chat speed for multi-model planning, minimal per-stage context, and a strict-manager retry that throws away a contaminated agent session instead of arguing with it.

What it is
#

A local web app (installed by curl, npm, Homebrew, Scoop, Bun, pnpm, yarn, or Docker) from looptroop-ai that runs on top of OpenCode. The pipeline has three layers: planning, where an adaptive interview produces a PRD that an LLM council drafts, anonymously scores, and refines into atomic “beads” (independent units with acceptance criteria and target files); execution, where OpenCode implements each bead in its own git worktree; and shipping, where a final verification pass hands off to your main branch. Its context engineering rebuilds minimal per-status context for each phase rather than passing a bloated transcript, and its Ralph-style retry preserves a compact error trace, discards the degraded workspace, and retries in a fresh worktree, which the site claims sustains ten or more hours of unattended execution. A human approval gate sits before any code is committed, and the README and docs cover the ticket state machine, prompts, API, and CLI.

Status
#

Active and early: about 162 stars and 14 forks as of 2026-10-15, created 2026-03-03, with commits through 2026-10-14 and a latest release of v0.6.1 (2026-10-13), a point release after v0.6.0 (2026-10-01), the major update (80 commits since v0.5.9) that added OpenCode v2 support, traceable skip reasons and decision history, and more than doubled the passing test count to 8,099.

Star History Chart

The documentation is unusually complete for an alpha, but the project’s own status page calls it “early alpha software” with bugs still likely, and it lists two configured constraints: councils of 2 to 10 models, and one active ticket per project in the execution band. A repo-scale orchestrator that can only run one ticket at a time per project is betting that depth beats throughput, which is the opposite of the parallel-session tools beside it.

Strengths
#

  • Multi-model council planning with independent drafting, anonymous voting, and winner refinement attacks the single-model bias problem directly.
  • Atomic bead decomposition with explicit acceptance criteria and per-bead isolation prevents cascading failures across a large change.
  • Minimal, rebuilt per-phase context is a principled answer to context rot in long autonomous runs.
  • Fresh-context recovery (the Ralph loop) is more robust than retrying inside a degraded session.
  • Free, MIT, 100% local, with a complete DOCUMENTED lifecycle and per-bead diff review before merge.

Cautions
#

  • Early alpha, 154 stars, single vendor, and one active ticket per project, so it is not a parallel fan-out tool.
  • OpenCode is the only execution backend; no Claude Code, Codex, or others.
  • Agents run with full local user privileges, and the docs themselves recommend running inside a disposable VM or cloud sandbox.
  • Planning is deliberately slow: the interview can take over an hour, which is a poor fit for small tasks.
  • The local GUI plus Node and OpenCode prerequisites make setup heavier than a CLI tool.

Pricing
#

Free and open source under MIT, with no telemetry. You pay only for the model providers OpenCode is configured to use.

Compared to
#

  • Crewplane: a CLI-first Markdown workflow runner; choose LoopTroop when you want council planning and a GUI rather than repo-owned workflow files.
  • Sandcastle: a TypeScript library that sandboxes and merges agents; choose LoopTroop when the problem is planning and recovery for big tickets, not sandbox plumbing.
  • Conductor: the polished Mac dashboard for many small parallel sessions; choose LoopTroop for fewer, long, high-correctness feature runs.

Bottom line
#

Recommended for engineers who want a local, MIT, planning-first orchestrator for large OpenCode tickets with human approval and strong recovery. Not for anyone who needs multi-harness support, parallel tickets, or a mature tool.

Changes
#

  • 2026-09-27 - Created.
  • 2026-10-02 - Recorded v0.6.0 (October 1, OpenCode v2 support, traceable skip reasons, and a doubled test suite) as the new latest release and refreshed counts.
  • 2026-10-07 - Added the looptroop-ai/LoopTroop star history chart to the Status section.

See also
#

References
#