Project Voltron

Specialized agent teams for Claude Code, served via MCP. Scaffold Unity, web, or general projects with battle-tested subagent templates — Alexandria-aware, self-improving, and auto-updating.

v2.4.1 MCP Server Unity Web / Fullstack 13 Agents MIT License

Docker Execution

The scrum-master launches each specialist agent inside a Docker container with --dangerously-skip-permissions via the run_agent_in_docker MCP tool. You run Claude Code normally on your desktop — Docker is handled behind the scenes.

🎯
Scrum-master creates a work plan Plans tasks, assigns agents, registers progress. Checks that Docker is available and Dockerfile.voltron exists.
🐳
Calls run_agent_in_docker for each task The tool builds the image (cached), mounts the project + OAuth credentials, composes the prompt from the agent template + CLAUDE.md + task description, and launches Claude Code inside the container.
Agent executes autonomously Inside Docker with --dangerously-skip-permissions — creates files, runs tests, fixes errors, iterates — with zero manual approval prompts. File changes are visible on the host via volume mount.
Output returned to scrum-master Scrum-master reviews the result, updates progress, and proceeds to the next task. Dashboard auto-refreshes every 5 seconds to reflect each transition.
Why Docker? Docker provides OS-level isolation from your host system, making --dangerously-skip-permissions safe to use. Agents can create files, run tests, install packages, and iterate on errors without any human gating — while your host stays protected.
No workflow change needed Run Claude Code normally on your desktop. The scrum-master handles Docker automatically — just ensure Docker is installed and running, and Dockerfile.voltron exists (generated by scaffold_project).

Progress Dashboard

Monitor agent work in real time. The scrum-master tracks every task, and a live HTML dashboard auto-refreshes every 5 seconds.

📊
Automatic tracking Scrum-master calls update_progress before and after each agent invocation — status, phase, and notes are recorded to .voltron/progress.json
🌐
Live dashboard Every progress update regenerates .voltron/dashboard.html — a self-contained page with phase timelines, status badges, and agent cards. Auto-opens in your browser on the first update.
💬
In-chat view Call get_progress at any time to see a formatted summary or detailed task table directly in the Claude Code chat window.
MCP Tool Purpose
update_progress Log a task status change (queued, in_progress, completed, failed, blocked). Auto-regenerates the HTML dashboard.
get_progress View a formatted progress summary or detailed task table in the chat window.
generate_dashboard Manually regenerate the standalone HTML dashboard at a custom path.

The Agent Teams

Scaffold the right team for your project type. Every project gets a scrum-master coordinator.

Core — all projects

Scrum Master @agent-scrum-master

Reads backlogs and project plans, breaks work into agent-sized tasks with dependencies and acceptance criteria. Never implements — only plans and delegates.

Project Planner @agent-project-planner

Researches tech stacks, designs architecture, defines data models and API contracts. Produces comprehensive project plans for scrum-master to decompose.

Researcher @agent-researcher

Deep research specialist. Finds any information — technical docs, APIs, pricing, competitors, legal text, community consensus — using web search, live page navigation, and structured extraction.

Unity

Scene Architect @agent-scene-architect

GameObject hierarchy, prefabs, scene composition, transforms, and components. Builds the world.

C# Dev @agent-csharp-dev

MonoBehaviours, ScriptableObjects, gameplay systems, and editor tools. Writes clean, performant Unity C#.

Shader Artist @agent-shader-artist

Shaders, materials, VFX Graph, post-processing, and render features. Knows URP, HDRP, and Built-in.

Build Validator @agent-build-validator

Console monitoring, compile checks, Play Mode smoke tests, pre-commit validation. Read-only watchdog.

Asset Manager @agent-asset-manager

Folder structure, import settings, naming conventions, and project audits. Keeps things organized.

Web / Fullstack

Fullstack Dev @agent-fullstack-dev

React/TypeScript frontend and Node.js/Express backend. Handles components, APIs, hooks, and data flow.

DevOps Engineer @agent-devops-engineer

Terraform, CI/CD pipelines, Docker, Fly.io, and AWS. Infrastructure as code and deployment automation.

UI Designer @agent-ui-designer

CSS architecture, responsive and mobile-first layouts, theming, PWA manifests, and accessibility.

QA Tester @agent-qa-tester

Vitest/Playwright tests, Lighthouse audits, bundle size analysis, and quality gates before merging.

Quick Start

Three steps from zero to a full agent team in your project.

1

Install

Clone and register as a global MCP server:

git clone https://github.com/7ports/project-voltron.git
cd project-voltron && npm install

claude mcp add --scope user \
  project-voltron -- \
  node /path/to/project-voltron/src/index.js
2

Scaffold

In any project, ask Claude Code:

"Scaffold this Unity project with Voltron agents"

# or for web projects:
"Scaffold this web project with Voltron agents"
3

Develop

Fill in CLAUDE.md, then invoke agents:

@agent-scrum-master  ← plan your sprint
@agent-csharp-dev    ← write Unity scripts
@agent-fullstack-dev ← build React + API
@agent-build-validator ← validate before commit

MCP Tools

9 tools for scaffolding, updating, and improving your agent team.

ToolDescription
list_templates List all templates, optionally filtered by project type
get_template Retrieve the full content of a specific template
scaffold_project Get the right agent set + auto-update hook for unity, web, fullstack, or general projects
get_auto_update_hook Get the .claude/settings.json hook that keeps agents current automatically — for projects scaffolded before v2.4.1
get_agent_usage_guide Usage guide for invoking and coordinating agents
check_for_updates Manually check if installed agent files are outdated vs. current templates
update_agent Get the latest content for a specific agent to replace an outdated file
submit_reflection Submit a post-session reflection on agent performance to drive improvements
list_reflections List stored reflections with pending improvement suggestions
run_agent_in_docker Launch a specialist agent inside a Docker container with full permissions — called by scrum-master
update_progress Update agent task progress — called by scrum-master before/after each agent invocation
get_progress View current agent task progress as a formatted dashboard in the chat window
generate_dashboard Generate a standalone HTML dashboard from progress data at .voltron/dashboard.html

Recommended Workflow

From backlog to shipped code, with agents coordinated at every step.

Scaffold Run scaffold_project to write CLAUDE.md, agent files, Dockerfile, launch script, and auto-update hook for your project type
📋
Configure Fill in CLAUDE.md with your project name, stack, conventions, and current work. Set up Docker execution.
🔬
Research & Design For new projects, invoke @agent-project-planner to research tech stacks, design architecture, and produce a comprehensive project plan
🎯
Plan Invoke @agent-scrum-master with the project plan — it decomposes it into agent-sized tasks with dependencies and acceptance criteria
🔨
Build Execute the plan — specialist agents look up setup guides in Alexandria before installing tools, and write findings back after
💡
Reflect Scrum-master automatically submits reflections at phase completion, blockers, and session end. Tool-specific discoveries are synced to Alexandria.

Self-Improving Agent Templates

Voltron gets better with every session through a built-in feedback loop.

Session runs with voltron agents scrum-master coordinates; specialists execute
Reflections submitted automatically scrum-master submits at phase completion, blockers, and session end — per-agent feedback with suggested changes
Reflections accumulate in reflections/ structured JSON, tied to project type and agents used
GitHub Actions runs Mon/Wed/Fri The reflection-processor agent (Sonnet 4.6) groups feedback by agent, prioritizes by frequency, applies targeted template edits, and opens a PR
PR reviewed and merged human reviews the diff before it reaches main — version bumped in package.json
Projects auto-update at next session start the UserPromptSubmit hook detects the version mismatch and silently rewrites all installed agent files in place — no manual steps needed

Alexandria Integration

Voltron agents are paired with Project Alexandria — a shared tooling knowledge base — so setup knowledge accumulates across every session.

scrum-master plans a new project calls get_project_setup_recommendations from Alexandria to surface required tools before work begins
specialist agents set up tools call quick_setup before installing anything — mandatory, not optional. Agents follow battle-tested guides rather than improvising from training data
agents discover platform-specific fixes call update_guide after completing integrations — workarounds, version notes, and gotchas flow into Alexandria
knowledge is available to every future agent the next project that touches the same tool gets accurate, up-to-date guidance automatically