ClosedLoop is an AI platform that brings the speed of individual AI-driven development to the full software development team. We're offering our agents as open sourced Claude Code plugins because we just couldn't keep this a secret for ourselves — check out our agents for planning, code reviews, judging quality and more that outperform Opus 4.6 and Sonnet 4.5 out of the box.
Bootstrap. Plan. Code. Ship. It's that simple.
LLMs are great at non-deterministic content generation — horrible at being repeatably correct.
That's why we took Claude Code and extended it with a lightweight multi-agent orchestration workflow paradigm that works for us; modeling how we collaborate as a team.
Optimized for efficiency & correctness to produce code that lands without the churn; it's grounded in your codebase and outperforms Opus 4.6 out of the box at half the cost.
What's more impactful is that it allowed our team of engineers to shift left; reviewing and approving sprints-worth of work scope in documented implementation plans and generating the code while we slept.
Tickets become Tasks. Epics become Features. Sections of your quarterly roadmap land in a few PRs.
Multi-repository, adaptive self-learning, & artifact-bound phased workflow gates that loop until correct.
Close the Loop on your SDLC with the same tools that made us 400% faster today.
- Python 3.11+ (3.13 recommended)
- jq
- Claude Code
One-line install — installs the five Symphony runtime plugins at user scope and keeps them auto-updated:
curl -fsSL https://raw.githubusercontent.com/closedloop-ai/claude-plugins/main/install.sh | bashThe installer installs code, code-review, judges, platform, and self-learning, then verifies those runtime plugins are present with existing install paths and enabled user-scoped entries. It re-enables disabled user-scoped runtime plugins and attempts to remove stale project-scoped duplicates when Claude reports a usable projectPath. If Claude reports a project-scoped entry without a usable project path, the installer prints the project-directory uninstall command and still repairs the user-scoped install. The bootstrap plugin remains available in the marketplace for manual installation, but it is not part of the default runtime install.
Or install interactively from within Claude Code:
claude /plugin marketplace install closedloopThis PR adds Codex metadata for the code-review plugin as a first
representative cross-harness install path:
codex plugin marketplace add https://github.com/closedloop-ai/claude-pluginsThen install ClosedLoop Code Review from the Codex Plugins UI. The same
.codex-plugin/plugin.json and .agents/plugins/marketplace.json pattern can
be repeated for the other ClosedLoop plugins.
Then start using the plugins:
# Plan. Code.
claude /code:code --prd requirements.md
# Review.
claude /code-review:startDevelopment setup
git clone git@github.com:closedloop-ai/claude-plugins.git
cd claude-plugins
git config core.hooksPath .githooks
See CONTRIBUTING.md for development setup, workflow, and code style guidelines.
The evals/code-review/ directory contains a small human-review eval set for
review finding quality, finding verification, and inline PR comment drafting.
The cases are harness-neutral so the same plugin behavior can be checked in
Claude Code, Codex, or another agent workspace before release.
If you publish plugins through Telvine, keep runtime telemetry metadata-only:
skill.invocation.start, skill.invocation.end, and skill.invocation.error
for skill behavior, plus plugin.component.invoked and
plugin.component.error for commands, agents, tools, or other non-skill
components. Do not emit prompts, repository file contents, connector payloads,
tool arguments, GitHub tokens, inline comment bodies, or model outputs.
Our claude code plugins are a low-key engineering preview of the agents that run the larger ClosedLoop platform. These agents should be used for testing in trusted environments.