Multi-agent autonomous building pipeline for Claude Code. Describe a goal, Claude Forge decomposes it into floors and ships real working code. Reimplemented from the original AskElira/AskElira3 Node prototype as native Claude Code sub-agents — no separate Hermes runtime, no Node infrastructure, no SQLite. The Claude Agent SDK is the runtime.
You say:
/claude-forge:build a REST API for a todo app with SQLite, JWT auth, and tests
Claude Forge:
- Elira decomposes the goal into 3-7 floors with dependencies
- For each floor (in dependency-aware parallel waves):
- Alba researches the floor (libraries, patterns, current docs)
- Vex Gate 1 validates the research is sufficient
- David writes real files into
./workspaces/<slug>/ - Syntax + smoke check (
node --check,python3 -m py_compile) - Vex Gate 2 validates the build
- Elira approves or rejects
- If rejected, Steven diagnoses + patches, loops back to Vex (max 5 iterations)
- After 5 failures, Rescue rebuilds with a simpler approach
- Reports LIVE / BLOCKED status per floor + writes a summary
All file writes happen through Claude Code's native tools (Write, Edit, Bash). Sub-agents are real Claude Code sub-agents in agents/. Slash commands are real Claude Code slash commands in commands/. No external runtime.
/plugin marketplace add AskElira/claude-forge
/plugin install claude-forge@claude-forge-marketplace
git clone https://github.com/AskElira/claude-forge .claude/plugins/claude-forgeThen reload Claude Code (/reload or restart). The plugin's agents, commands, and skill will appear automatically.
Drop the cloned folder into ~/.claude/plugins/claude-forge/ to use across all your projects.
/claude-forge:build a Python CLI that scrapes Hacker News top stories and emails them daily
/claude-forge:plan a Next.js dashboard with Supabase auth and Stripe billing
/claude-forge:fix workspaces/my-goal-slug 3 "the API returns 500 on /todos POST"
/claude-forge:status
Goal
│
▼
Elira plans floors ──► 3-7 floors with dependsOn
│
▼
Build waves (parallel where deps allow)
│
▼ per floor (max 5 iterations):
│
│ Alba research ─► Vex Gate 1 ─► David build
│ │
│ ▼
│ Syntax + smoke check
│ │
│ ▼
│ Vex Gate 2
│ │
│ ▼
│ Elira approve?
│ ┌──┴──┐
│ YES NO
│ │ │
│ LIVE Steven patches ─► loop
│ │
│ (after 5)
│ ▼
│ Rescue rebuilds
│ ▼
│ Elira final review
│ │
│ LIVE or BLOCKED
claude-forge/
├── .claude-plugin/
│ ├── plugin.json # plugin metadata
│ └── marketplace.json # for git-based marketplace install
├── agents/
│ ├── elira.md # architect, planner, approver
│ ├── steven.md # fixer, debugger
│ ├── alba.md # researcher
│ ├── david.md # builder (writes real files)
│ ├── vex.md # validator (two gates)
│ └── rescue.md # fresh-perspective rebuilder
├── commands/
│ ├── build.md # /claude-forge:build — full pipeline
│ ├── plan.md # /claude-forge:plan — preview floors
│ ├── fix.md # /claude-forge:fix — manual Steven
│ └── status.md # /claude-forge:status — list workspaces
├── skills/
│ └── claude-forge-build/
│ └── SKILL.md # auto-trigger when user asks to "build"
├── examples/
│ └── todo-api.md # sample run-through
├── CLAUDE.md # project guidelines (loaded into context)
├── LICENSE
└── README.md
| Agent | Role | Model |
|---|---|---|
| elira | Plans goals → floors. Approves or rejects builds. | opus |
| steven | Diagnoses failures. Writes patches. Runs setup. | opus |
| alba | Researches floor requirements + library APIs. | sonnet |
| david | Writes real, complete, runnable code files. | sonnet |
| vex | Validates research (Gate 1) and builds (Gate 2). | sonnet |
| rescue | Rebuilds from scratch after 5 failed iterations. | opus |
You can also call any agent directly with the Task tool:
Use the alba sub-agent to research GraphQL subscriptions in Apollo Server v4.
| AskElira3 (Node app) | Claude Forge (Claude plugin) |
|---|---|
| Hermes is a custom LLM orchestrator | Claude Code IS the orchestrator |
| SQLite stores goals/floors/logs | Workspace + JSON files store state |
| Express dashboard at localhost:3000 | Use Claude Code's built-in UI / chat |
| Custom LLM client with retries | Claude Code handles model calls |
| Circuit breakers, watchdogs | Claude Code's existing timeouts |
| Agents are JS modules | Agents are Markdown sub-agent files |
npm start && askelira3 "<goal>" |
/claude-forge:build <goal> |
The mechanism is identical. The infrastructure is gone — Claude Code already provides everything Hermes had to build from scratch.
These are enforced by every agent's system prompt. They are not suggestions.
- No stubs. No TODO, FIXME, placeholder. Every file ships complete and runnable.
- JSON-only outputs from sub-agents. One document per response.
- Workspace boundary. No agent writes outside its goal's workspace.
- Approval is final. Elira's verdict ends the iteration loop.
- Bias toward passing. Vex blocks only on real blocking issues.
- 5-iteration cap per floor, then Rescue, then BLOCKED.
MIT. See LICENSE.
Reimplemented from AskElira/AskElira3. Inspired by forrestchang/andrej-karpathy-skills.