Open Source · Claude Code Plugins

Pipelines that
think in parallel.

A marketplace of multi-agent plugins for Claude Code. Autonomous development, review, and research — orchestrated as a single slash command.

12
Plugins
~150
Specialised agents
MIT
Licensed
01 — What it is

One command fans out
into a swarm of agents.

Each plugin is a marketplace you pin like any other. Behind every slash command sits a pipeline of specialised agents that research, plan, write, review, and test — coordinated through phased stages with human approval gates. You stay the manager; the agents do the fan-out.

02 — How the swarm behaves
/ fan-out

Parallel by default

Independent work units run simultaneously across many agents — up to eight coders at once on a single build.

/ gate

Approval gates

No code changes land without your sign-off on the plan. The pipeline pauses and waits for you.

/ converge

Iterative convergence

Fix → verify → review loops run with bounded retries until the criteria actually hold.

/ disclose

Progressive disclosure

Orchestrators load each phase on demand, keeping token cost low even on deep, multi-stage runs.

03 — The pipeline

From prompt to merged,
one phase at a time.

01
/build
Kick off the run
02
Research
Explore the codebase
03
Plan
Approval gate
04
Code ×N
Parallel workers
05
Review
5-lens checks
06
Test
Verify & ship
04 — The marketplace

Twelve plugins. Pick what you need.

Each plugin is a self-contained pipeline you install on its own. They compose — most real tasks chain commands across three or four of them.

mz-dev-pipeFlagship

Autonomous dev pipelines — plan, code, review, test, in parallel.

/build/audit/debug
mz-dev-git

Multi-lens deep review for local branches and GitHub pull requests.

/review-branch/github-review-pr
mz-research-pipe

Research & content pipelines that produce reports, not code changes.

/deep-research/explain
mz-creative

Multi-perspective lens panels plus naturalized writing pipelines.

/brainstorm/expert
mz-gov

Turn a big decision into a reviewed ADR / RFD behind a human sign-off.

/govern/gov-review
mz-knowledge

Full lifecycle for a personal Obsidian vault — capture to grounded Q&A.

/vault-ingest/vault-answer
mz-biz-outreach

Lead-gen pipeline plus a dossier-card lifecycle for every company.

/outreach-research
mz-job-outreach

CV-to-vacancies and freelance-gig discovery with grounded cover letters.

/job-search/freelance-search
mz-design

Design specs refined by four critics behind a WCAG 2.2 AA hard gate.

/design-document
mz-dev-base

Standalone agents and 14 curated coding rules for everyday work.

/init-rules/construct-skill
mz-memory

Cross-session project memory that persists knowledge automatically.

/memory-note
mz-funny

Evidence-anchored code roasting in seven character voices.

/do-roast

Get it running

You need Claude Code on a recent version (run claude --version and update if /plugin isn't recognised) and Claude access — a Pro, Max, Team, or Enterprise plan, an Anthropic Console key, or a Bedrock / Vertex provider. The free tier doesn't include Claude Code.

Add the catalogue. One command pins the whole marketplace: /plugin marketplace add DoctorMozg/claude-pipelines

Install a plugin. Start with the flagship; add others the same way: /plugin install mz-dev-pipe@mozg-pipelines

Run it. Invoke a slash command and the pipeline takes over. Plugins update in place from the marketplace: /build

A real run, end to end

Say you're adding a feature you've never built before. Start with /deep-research to map the approach and the trade-offs — a report, no code yet. Hand the findings to /build: it plans, you approve the plan, then up to eight coders work the implementation in parallel. Run /review-branch to put the result through multiple lenses before it goes anywhere. If the decision was load-bearing, /govern turns it into a reviewed ADR. You stay the manager at every gate. What lands is a reviewed branch with an audit trail — not just generated code.

Won't a swarm burn my tokens?

Parallel agents cost more than a single prompt — there's no magic that hides that. What the design avoids is the runaway loop. Orchestrators load each phase on demand, so context doesn't balloon as a run goes deep. The fix → verify → review loops stop at a cap instead of grinding forever. And you choose the scope: run one plugin and one command for a small job, and reserve the full eight-coder fan-out for work that's genuinely multi-step. The trade you're making is wall-clock time and review depth, not a smaller bill on every task.

Where your code and model traffic go

The plugins add no hosted backend and no telemetry of their own. They're skills, agents, and commands that run inside your local Claude Code session under MIT — there's no SGON.AI server in the path. Where inference goes is Claude Code's decision, and you control it: the public Anthropic API by default, or Amazon Bedrock, Google Vertex, or an OpenAI-compatible gateway via environment variables. So a privacy-sensitive team can run the full pipeline while keeping model calls inside a cloud account and region they control — useful under the UAE PDPL (Federal Decree-Law 45/2021) when cross-border transfer is the concern. Be clear on the limit: these plugins are orchestration, not a residency or compliance product. Residency comes from how you route Claude Code, and the tooling touches your source code, not customer data — unless your code or tests contain it. Pair it with orno, the sibling project, to bound and record agent runs in CI.

Frequently asked questions

Do I need an Anthropic API key?

You need Claude access — a Claude Code subscription (Pro, Max, Team, or Enterprise), an Anthropic Console key, or a Bedrock / Vertex provider. The plugins don't add their own auth.

Can I run it without the public cloud, against my own endpoint?

Yes. Point Claude Code at Bedrock, Vertex, or an OpenAI-compatible gateway through environment variables, and the pipelines follow — they're orchestration on top of whatever provider Claude Code uses.

Will it push or merge code without my approval?

No code is written until you approve the plan — the pipeline pauses at that gate and waits. The review plugins surface findings on a branch; they don't auto-merge.

Is it really free?

The pipeline source is MIT and free. You pay only for the model tokens your runs consume, billed by whichever provider you've configured.

When is this overkill?

For a one-line fix, skip the swarm. Reach for it on multi-step features, deep reviews, and research, where the parallelism and the gates earn their keep.

Built in the open

Read every line.
Run it yourself.

Nothing hidden, nothing locked. The full pipeline source ships under MIT — clone it, audit it, extend it.

View on GitHub
MIT · claude-pipelines