claude-code-teams

Multi-agent feature pipeline for your projects, built on Claude Code Agent Teams. One prompt → legacy analysis → plan → parallel FE/BE build → parallel QA/security/perf review → PR → cleanup — fully autonomous, only pausing for genuine business decisions and the final merge.

[!IMPORTANT] This is a Claude Code plugin, not an npm library — npm install claude-code-teams will not install it into your project. This package is published to npm only so the name is reserved and the plugin is discoverable in npm search; the actual install mechanism is Claude Code's own /plugin command, pulling from the GitHub repo, not from the npm registry. Skip straight to Install below.

Source and marketplace manifest: github.com/iammuhammadasimofficial/claude-code-teams (this folder, plugins/claude-code-teams/, is one plugin inside that repo's marketplace — the marketplace manifest itself lives at the repo root).

Requirements

• Claude Code, updated to a recent version (Agent Teams is required):

  claude update

• git on your PATH.
• Node.js 18+ (only needed if you run the helper scripts in scripts/).
• Agent Teams enabled — see step 2.

Install (step by step)

The install flow is the same on Windows, macOS and Linux. Only the clone folder and shell differ. Do it once per machine.

1. Pick a folder and clone the repo

Clone the repo somewhere permanent (you'll keep the clone around and git pull it for updates — don't delete it after installing).

Windows (PowerShell):

cd $HOME
git clone https://github.com/iammuhammadasimofficial/claude-code-teams.git
# → clone lands at C:\Users\<you>\claude-code-teams

macOS / Linux (bash/zsh):

cd ~
git clone https://github.com/iammuhammadasimofficial/claude-code-teams.git
# → clone lands at ~/claude-code-teams

The plugin itself lives at claude-code-teams/plugins/claude-code-teams/, but the marketplace manifest is at the repo root. Remember the full path to the claude-code-teams clone — you'll point Claude Code at it (not the plugin subfolder) in step 3.

2. Enable agent teams

Add the experimental flag to your user settings file:

• Windows: C:\Users\<you>\.claude\settings.json
• macOS / Linux: ~/.claude/settings.json

{
  "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }
}

If the file already has an "env" block, just add the key inside it.

3. Register the local marketplace and install the plugin

Open Claude Code (in any project), then run these slash commands inside the session. Point the marketplace at the repo root (the directory that contains .claude-plugin/marketplace.json) — not the plugin subfolder.

Use forward slashes / in the path on every OS — including Windows. Give the full absolute path to the claude-code-teams clone.

Windows (adjust <you>):

/plugin marketplace add C:/Users/<you>/claude-code-teams
/plugin install claude-code-teams@shipyard

macOS / Linux:

/plugin marketplace add ~/claude-code-teams
/plugin install claude-code-teams@shipyard

• shipyard is the marketplace name (from marketplace.json), claude-code-teams is the plugin name.
• Tip: if you launched Claude Code from inside the clone, you can use a relative path instead: /plugin marketplace add .
• Tip: you can skip the manual clone and add the GitHub repo directly instead: /plugin marketplace add iammuhammadasimofficial/claude-code-teams.

4. Reload so the plugin activates

/reload-plugins

No restart needed. Verify it loaded:

/claude-code-teams:help

You should see the command/skill reference. Done.

Updating to new changes

When new changes land in the repo, update your existing clone — do not re-clone into a new folder.

Windows (PowerShell):

cd $HOME\claude-code-teams
git pull

macOS / Linux:

cd ~/claude-code-teams
git pull

Then, inside a Claude Code session, refresh the marketplace metadata and reload:

/plugin marketplace update shipyard
/reload-plugins

If a reload alone doesn't pick up the new version, reinstall:

/plugin install claude-code-teams@shipyard
/reload-plugins

Reloading plugins

To re-apply plugin changes in a running session at any time (after an update, or after editing a skill/agent locally):

/reload-plugins

This reloads all active plugins without restarting Claude Code.

Permissions (no "permission denied" on any OS)

The shell scripts used by this plugin's hooks are committed with the executable bit set, so hooks run cleanly on macOS and Linux out of the box. Windows ignores the exec bit and is unaffected.

If a macOS/Linux user ever sees a permission denied from a hook, restore the bit on the clone:

chmod +x ~/claude-code-teams/plugins/claude-code-teams/scripts/*.sh

Running a feature

[!IMPORTANT] Launch it right, or it won't be autonomous. Open your Claude Code terminal in the root folder of the target project (not a subfolder, not this plugin repo), and start it in bypass-permissions mode — e.g. claude --dangerously-skip-permissions, or set "permissions": { "defaultMode": "bypassPermissions" } in the project's .claude/settings.json. Bypass mode lets the team run its own commands (git, npm, playwright, worktrees, scripts) unattended instead of pausing on per-tool permission prompts. Only use bypass mode in repos you trust.

With the terminal opened at the project root in bypass-permissions mode:

/claude-code-teams:orchestrate-feature Policy Renewals module — legacy source ../blanka, FRD in ClickUp task ABC-123

Or just describe the feature in natural language; the lead follows the same pipeline and runs it hands-off to an open PR. The final PR merge stays with you; everything up to the open PR is automatic. Quality gates (≥90%/100% coverage, zero blockers, tests green) are enforced by looping, not by asking you.

Models are pre-configured per agent and deliberately tiered to keep weekly Opus usage low without dropping quality: Opus for the 5 judgment-critical agents only, Sonnet for all implementation and analysis, Haiku for mechanical plumbing. See Agent roster for the full table. Override in an agent's frontmatter if you want.

First time in a repo: run /claude-code-teams:agent-team-orc once to pick a roster and save the team (.claude/team.json, commit it). After that every orchestrate-feature/review-pr uses it.

Commands

Type these in Claude Code. Run any command bare (no args) for a guided wizard; natural language ("list my teams", "review PR 142") works too.

| Command | What it does | |---|---| | /claude-code-teams:agent-team-orc | Team wizard — create / list / show / rename / delete a team for this project | | /claude-code-teams:orchestrate-feature team <name>: <feature + docs> | Full autonomous pipeline: BA/legacy → plan → parallel build → test wave (90%/100% coverage + Playwright) → QA/security/perf reviews → PR → cleanup. Only stops for genuine business decisions and the final merge | | /claude-code-teams:review-pr team <name>: <pr#> [module] | Multi-agent PR review with KB parity; consolidated BLOCKER/MAJOR/MINOR verdict | | /claude-code-teams:figma-to-ui <figma-url> | Figma design → implemented UI. Pulls design context/tokens via the Figma MCP, maps them onto the project's existing design system, implements per framework, then verifies with Playwright screenshots against the design and reports deliberate deviations | | /claude-code-teams:axiom-triage [dataset] [window] | Pulls real production errors from an Axiom dataset, groups them by normalised signature, ranks by user impact and regression-since-deploy, root-causes each group, and fixes with a failing regression test first | | /claude-code-teams:kb [module] | Build the full product knowledge base and its browsable /kb viewer, in the project's own stack (not a separate docs site). Bare = full build; with a module name = add or refresh just that module. Sources from knowledge-base/, FRDs and the code; writes Markdown to .claude/knowledgebase/; new pages appear with no code change. /kb is absent in production, open elsewhere | | /claude-code-teams:status | Running teams/agents + task progress (read-only, terminal) | | /claude-code-teams:help | Full command/skill/script/hook reference |

Skills (239 — model-invoked; force one with /claude-code-teams:<skill>)

The lead auto-loads whichever skills fit the task. All skill content is vendored verbatim from upstream sources (see Skill provenance) — none of it is hand-written.

| Area | Skills | |---|---| | Next.js / React | nextjs-app-router-patterns, nextjs-typescript, next-dev-loop, next-cache-components-optimizer, react-best-practices, react-state-management, tailwind-design-system, composition-patterns | | Angular | 12 skills — angular-developer, angular-signals, angular-forms, angular-routing, angular-http, angular-di, angular-directives, angular-ssr, angular-testing, angular-tooling, angular-component, angular-new-app | | Vue / Nuxt | 13 — vue-best-practices, vue-debug-guides, vue-pinia/router/testing/options-api/jsx-best-practices, vue-adaptable-composable, vueuse-functions, pinia, nuxt, nuxt-ui, nitro | | Svelte / Astro | svelte-core-bestpractices, svelte-code-writer, astro-framework, astro-learning | | React Native / Expo | react-native-best-practices (+tv, +brownfield), expo-* (11), eas-* (4), react-navigation, upgrading-react-native, assess-react-native-migration, create-react-native-library | | Flutter / Dart | 22 — flutter-* (10) and dart-* (12) | | Databases | mongodb-* (3), redis-* (3), postgresql-optimization, postgresql-code-review, postgres-operations, mysql-operations, vitess-scaling, mariadb-* (6), mysql-to-mariadb, oracle-to-mariadb, sql-optimization, sql-code-review, database-migration | | Cloud | AWS (launch-with-aws, aws-* ×6, amazon-* ×3, rds-oss, querying-aws-cloudwatch), Azure (azure-* ×10), GCP (cloud-run-basics, cloud-sql-basics, alloydb-basics, gke-cluster-creation, google-cloud-* ×4, cloud-logging-configuration-basics) | | IaC / CI / shell | terraform-* (5), multi-stage-dockerfile, github-actions-hardening, github-actions-efficiency, bash-pro, linux-shell-scripting, powershell-windows | | Observability | axiom-sre, axiom-alerting, axiom-dashboards, axiom-query-metrics, axiom-metrics-chart, axiom-cost-control, axiom-spl-to-apl, axiom-writing-evals, sentry-security-review | | SEO / web quality | seo-audit, seo-technical, seo-schema-markup, seo-content-strategy, ai-seo, programmatic-seo, core-web-vitals, web-performance, web-accessibility, web-quality-audit, web-quality-best-practices, web-design-guidelines, responsive-design, design-system-patterns, frontend-design | | Build tooling | vite, vitest, unocss, vitepress, tsdown, pnpm, turborepo | | Legacy / backend | laravel-* (5), php-best-practices, nestjs-best-practices, better-auth-* (4), architecture-patterns, api-design-principles, codebase-design, domain-modeling | | Integrations | paymob-integration, webengage-integration, liteapi-integration | | Security | security-audit, sast-configuration, gha-security-review, pci-compliance, threat-mitigation-mapping | | Process | kb-builder, product-kb, plan-lifecycle, writing-plans, executing-plans, to-spec, to-tickets, frd-builder, commit, pr-writer, clickup-workflows, git-worktree-discipline |

Full list: /claude-code-teams:help, or ls skills/.

MCP servers (.mcp.json)

| Server | URL | Used by | |---|---|---| | clickup | https://mcp.clickup.com/mcp | FRDs/tasks, comments, attachments; clickup-ops | | figma | https://mcp.figma.com/mcp | design context for figma-to-ui, prototype-builder, frontend-dev | | axiom | https://mcp.axiom.co/mcp | error datasets for axiom-triage, observability/incident agents |

Figma and Axiom authenticate over OAuth — the first call in a session opens a browser authorization. ClickUp additionally needs CLICKUP_API_TOKEN (see step 2). If a server isn't connected, the dependent command says so and offers a fallback rather than inventing data.

Figma's official skills (install alongside)

Figma publishes its own Claude Code plugin with 12 first-party skills for driving the Figma MCP — figma-use (which Figma marks mandatory before any use_figma call), figma-design-to-code, figma-generate-design, figma-generate-library, figma-code-connect, figma-implement-motion, plus FigJam/Slides/diagram skills.

Our marketplace lists it, so it installs from Figma's repo in one line:

/plugin install figma@shipyard
/reload-plugins

It is referenced, not vendored — deliberately. Figma's repo ships no LICENSE file, so copying its files into this repo would be an unlicensed redistribution; and because the skills track the MCP server's own version (currently 2.2.81), pointing at source keeps them current instead of frozen at a sync date. figma-to-ui detects them and defers to them for MCP mechanics, handling only the project-specific part: mapping onto the project's design system, picking the right agent, and checking its quality gates.

Scripts (scripts/, run with node/bash)

| Script | Purpose | |---|---| | node scripts/teams-monitor.mjs [--watch\|--json] | Terminal view of running teams | | bash scripts/worktree.sh new <feature> [base] \| list \| remove <feature> | Worktree + unique dev port | | bash scripts/coverage-gate.sh [base-branch] | Blocks PR below 90% coverage | | node scripts/kb-index.mjs [--check] | Rebuild the /kb nav manifest; --check fails on drift (CI gate) | | node scripts/kb-screenshots.mjs [baseURL] | Playwright screenshots for the product KB |

Hooks (always on)

• worktree-guard — blocks code edits on main/develop in the primary checkout (config/docs exempt).
• task-gate — blocks task completion if lint/typecheck fail.
• kb-sync — refreshes the /kb nav manifest whenever a file under .claude/knowledgebase/ is written, so the KB can never drift behind its content. Silent unless it changed; never blocks a tool call.

Monitoring

• Default (all OS): in-process panel — ↑/↓ selects a teammate, Enter opens its transcript, Esc interrupts, Ctrl+T toggles the task list.
• macOS/Linux optional split panes: install tmux, then claude --teammate-mode auto (or set "teammateMode": "auto"). Not supported in Windows Terminal — Windows devs use in-process mode or WSL+tmux.

Layout expected in each project repo

knowledge-base/<module>/   # parity contract (kb-builder skill format)
plans/upcoming|in-progress|done/
CLAUDE.md                  # project conventions — teammates load it automatically

Notes

• Agent Teams is experimental: if the terminal dies, in-process teammates are lost. All state (KB, plans, tasks) lives on disk, so re-run the skill and continue.
• Token cost scales per teammate. The pipeline runs phases sequentially and only parallelizes where it pays (FE+BE, the three reviewers).
• One feature = one worktree = one team. Multiple features → multiple lead sessions.

Agent roster (60)

The lead spawns only what the feature needs — 3–5 active teammates max. A specialist replaces a generalist in the wave; it doesn't stack on top. Adding agents to the plugin costs nothing until one is spawned.

Core pipeline (13) — ba-analyst, legacy-analyst, planner, prototype-builder, frontend-dev, backend-dev, integrations-dev, unit-test-engineer, e2e-playwright, qa-reviewer, security-reviewer, performance-reviewer, devops-vercel.

Specialists (47), spawned on demand:

| Area | Agents | |---|---| | Database | database-architect, database-optimizer, sql-pro | | Mobile | mobile-developer, flutter-expert, ios-developer | | Cloud / infra | cloud-architect, kubernetes-architect, terraform-specialist, deployment-engineer, devops-troubleshooter, observability-engineer, incident-responder | | Data / AI | data-engineer, data-scientist, ai-engineer, ml-engineer | | Quality | code-reviewer, architect-review, debugger, error-detective, test-automator, tdd-orchestrator, accessibility-expert | | Docs / design | api-documenter, docs-architect, ui-ux-designer | | Languages | typescript-pro, javascript-pro, python-pro, php-pro, golang-pro, java-pro, csharp-pro, rust-pro, shell-pro | | Frameworks | django-pro, fastapi-pro, graphql-architect | | SEO / growth | seo-structure-architect, seo-keyword-strategist, seo-meta-optimizer, seo-content-writer, seo-content-auditor, content-marketer | | Ops (mechanical) | git-ops, clickup-ops |

database-architect/database-optimizer/sql-pro and backend-dev have a documented handoff in both directions: backend-dev sends access patterns and slow-query plans; the database agents return the schema, index plan and migration steps, and never edit application code themselves.

Model and effort tiering

Tiered to hold weekly Opus usage down without dropping quality. Opus is 5 of 60 agents — reserved for work where a wrong judgement is expensive and the volume is low. Everything that scales with feature size runs Sonnet. Plumbing runs Haiku.

| Tier | Count | Agents | |---|---|---| | Opus / high | 5 | planner, qa-reviewer, security-reviewer, architect-review, incident-responder | | Sonnet / high | 43 | all implementation, analysis, database, cloud, mobile, data/AI, quality and language agents | | Sonnet / medium | 8 | seo-keyword-strategist, seo-meta-optimizer, seo-content-writer, seo-content-auditor, content-marketer, api-documenter, docs-architect, shell-pro | | Sonnet / low | 1 | prototype-builder (throwaway output) | | Haiku / low | 3 | git-ops, clickup-ops, devops-vercel |

Send mechanical work to Haiku. git-ops handles staging, commit messages, branch creation, pushes, PR opening and worktree setup/teardown. clickup-ops handles task reads/updates, status transitions, comments and screenshot/attachment uploads. Neither has design authority — both have explicit escalation limits (merge conflicts, history rewrites, failing gates, ambiguous requirements, destructive operations) and hand back rather than guessing. Keeping this traffic off Sonnet/Opus is where most of the token saving comes from, because it is the highest-volume work in a run.

Skill provenance

Every skill is vendored verbatim from an upstream source with a <!-- Vendored from … --> comment and a license: field in its frontmatter. Nothing is hand-written, so upstream fixes can be re-synced. Sources include the official first-party repos for Angular, Vue, Nuxt, Svelte, Expo, Flutter, Dart, MariaDB, PlanetScale, AWS, Microsoft/Azure, Google Cloud, HashiCorp, GitHub and Axiom, plus wshobson/agents for the agent definitions.

Licenses are mixed and attributed per file — check the frontmatter before reusing a skill outside this repo:

| License | Applies to | |---|---| | MIT | most skills; all agents (wshobson/agents) | | Apache-2.0 | AWS (aws/agent-toolkit-for-aws), GCP (google/skills) | | BSD-3-Clause | Flutter (flutter/agent-plugins), Dart (dart-lang/skills) — clause 3 forbids implying Flutter/Google endorsement | | MPL-2.0 | Terraform (hashicorp/agent-skills) — weak file-level copyleft; keep modifications to those files MPL and retain notices |

One caveat worth knowing: vercel-labs/agent-skills ships no LICENSE file — MIT is asserted only in its README and in some of its skills' own frontmatter. We record those as MIT on that basis (vercel-optimize, web-design-guidelines, react-best-practices, composition-patterns, nextjs-typescript). Every other source has a real LICENSE file, verified at vendoring time.

Re-syncing a skill means re-copying from upstream and updating the synced date — review the upstream diff first. The test suite enforces that every vendored skill keeps a license: field and a Vendored from … — synced <date> comment, so attribution can't silently rot.