Kiln

Kiln is a native plugin designed for Claude Code, providing a fully autonomous build pipeline from requirement dialogue to deliverable software. Its 7-step process (Onboarding → Brainstorm → Research → Architecture → Build → Validate → Report) is driven by 34 named agents, running fully autonomously after Step 3.

Core Capabilities#

• JIT Scope Partitioning: KRS-One splits build chunks from live codebase state rather than executing stale upfront plans
• Persistent Team Context: Cross-iteration memory via Claude Code native TeamCreate/SendMessage primitives
• 3-Layer Review Mechanism: Pair review → Judge Dredd dual-model QA Tribunal (anonymous reconciliation) → Argus end-to-end validation
• Built-in TDD Cycles: RED/GREEN/REFACTOR embedded in build loops
• Crash Recovery: .kiln/STATE.md captures pipeline position, /kiln-fire resumes from interruption
• Deep Brainstorming: Adapted from BMAD Method — 62 techniques + 10 categories + 50 facilitation methods, outputting VISION.md

Multi-Model Orchestration#

The primary model runs on Claude. With Codex CLI installed, GPT-series models can be dispatched for independent planning and dual-model anonymous review reconciliation, reducing single-model bias. Without Codex CLI, falls back to pure Claude single-model mode.

Architecture#

Implemented as Markdown prompts + Shell/Python hooks with no runtime daemon, loaded purely through the Claude Code plugin mechanism. The plugin directory contains 34 agent prompt definitions, 2 slash commands, lifecycle hooks (enforce-pipeline.sh for step ordering, task-dag-guard.py for preventing malformed task transitions, validate-state.py for state validation), 2 composable skills, and MCP configuration.

Worker lifecycle gating: REQUEST_WORKERS → REQUEST_WORKERS_READY → CYCLE_WORKERS (deterministic SubagentStart readiness gate) → WORKERS_SPAWNED.

Installation & Usage#

Prerequisites: Claude Code (npm i -g @anthropic-ai/claude-code), Node 18+, jq, Git identity config, environment variable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 (required). Optional: Codex CLI (dual-model mode), uvx (Fetch MCP), Playwright (E2E validation).

# Install via Claude Code plugin
claude plugin marketplace add Fredasterehub/kiln
claude plugin install kiln

# Install via shell script
curl -fsSL https://raw.githubusercontent.com/Fredasterehub/kiln/main/install.sh | bash

# Pre-check and launch
/kiln-doctor --fix
/kiln-fire

Updates require manual claude plugin update kiln@kiln. Test coverage includes lint fixtures and 19 scenario integration tests, all passing.

Use Cases#

• End-to-end software development (from requirements to tested, committed code)
• Multi-model collaborative code building (Claude + GPT dual-model review reconciliation)
• Large project iterative development (cross-iteration context accumulation)
• Brownfield project modernization (auto-detect and map existing codebase structure)
• Quality assurance pipeline (3-layer review + TDD + E2E validation)

Known Limitations#

• Depends on Claude Code experimental Agent Teams feature; future API changes may cause compatibility issues
• Model version references in README (Opus 4.7, GPT-5.5, etc.) may be hypothetical; actual correspondence unconfirmed
• No public production usage reports available
• Project marked as WORK IN PROGRESS; long-term maintenance plans unclear
• Author identity and background not disclosed