AI Architect

Findings → PRD → verified PR. An autonomous software-engineering pipeline for Claude Code that runs on real evidence: 17 codebase intelligence tools, 5 science-backed git analytics, 64 deterministic verification rules, and zero LLM-judges-LLM. Every claim traces to a paper. Every PR passes its own gates.

/plugin marketplace add cdeust/ai-architect-mcp
/plugin install ai-architect

Then run /ai-architect:run-pipeline from any project. Open source. MIT licensed.

What you get

Three MCP servers, one pipeline, one philosophy.

ai-architect — pipeline orchestration (49 MCP tools)

• 11 stages, 0–10: health → discovery → impact → integration → PRD → interview → review → implementation → verification → benchmark → PR

• 64 deterministic HOR rules across 10 categories (structural, security, resilience, observability, …)

• 5 verification algorithms — Chain of Verification, KS Adaptive Stability, Multi-Agent Debate, NLI Entailment, Zero-LLM Graph Verification

• 5 prompting algorithms — TRM Self-Refine, Adaptive Expansion (ToT/GoT), Metacognitive Monitor, Signal-Aware Thought Buffer, Collaborative Inference

• 10-dimension Plan Interview gate — deterministic PRD quality check with zero LLM calls

ai-codebase-intelligence — typed graph engine (17 MCP tools)

• Tree-sitter parsing for 13 languages (Python, TS, Swift, Go, Rust, Java, Kotlin, C#, Ruby, PHP, C, C++, JavaScript)

• Typed Pydantic API — GraphNode, GraphRelationship, 29 node labels, 19 relationship types, no raw dicts

• Hybrid search — BM25 full-text + Reciprocal Rank Fusion + process-aware ranking

• 5 science-backed git analytics:

  Tool	Paper	What it does
  ownership	Bird et al. 2011 (ESEC/FSE)	per-file ownership ratio + minor contributor count
  bus_factor	Avelino et al. 2016 (ICPC)	iterative top-contributor removal until orphan threshold
  churn	Nagappan & Ball 2005 (ICSE)	relative code churn metrics, defect predictor
  cochange	Gall 1998 + Zimmermann 2005	hidden coupling via association rule mining on commits
  dead_code	Grove 1997 + Tip 1999	conservative CHA-based reachability with confidence levels

cortex (optional) — persistent cognitive memory

Recall past findings before each pipeline run. Remember decisions after. The pipeline learns across sessions instead of starting fresh every time. See Cortex.

Related MCP server: Claude PM MCP

30-second start

git clone https://github.com/cdeust/ai-architect-mcp.git
cd ai-architect-mcp
./scripts/setup.sh

That installs both MCP servers, symlinks the 13 stage skills into ~/.claude/skills/, and registers the slash commands. Restart Claude Code, then:

/ai-architect:run-pipeline

That's it. The pipeline detects findings, scores them, generates PRDs, implements them, verifies, and opens PRs.

Real example — verified end-to-end

This repo dogfoods itself. PR #2 was produced entirely through MCP tools:

CI runs the same 544-test suite on every PR. No PR ships without green tests.

How it works

Findings (research papers, bug reports, ideas)
   │
   ▼
Stage 0  Health        Validate all 49 + 17 MCP tools
Stage 1  Discovery     Query codebase intelligence + churn/cochange + cortex recall
Stage 2  Impact        Compound scoring + ownership + bus_factor risk
Stage 3  Integration   Cochange coupling + dead_code cleanup + port design
Stage 4  PRD           Generate via ai-prd-generator (delegated)
Stage 4.5 Interview    10-dimension deterministic gate
Stage 5  Review        7 verification algorithms, loop on score < 0.85
Stage 6  Implementation One worker per file, dependency-ordered
Stage 7  Verification  64 HOR rules + build gate, loop on failure
Stage 8  Benchmark     Performance gates from prd-tests.md
Stage 9  Deployment    Full test suite, zero failures required
Stage 10 PR            Push, open PR with audit trail, save lessons to cortex

Three layers, strictly separated:

• Skills (skills/*/SKILL.md) define WHAT each stage does. Markdown only, zero code.

• Tools (mcp/, ai-codebase-intelligence/) define HOW. Python only, zero markdown logic.

• Claude decides WHY. Reads the skill, calls tools, evaluates output, retries or proceeds.

Installation

Plugin (Claude Code)

/plugin marketplace add cdeust/ai-architect-mcp
/plugin install ai-architect

Manual (Claude Code, Cursor, VS Code, Claude Desktop)

git clone https://github.com/cdeust/ai-architect-mcp.git
cd ai-architect-mcp
./scripts/setup.sh

setup.sh installs both MCP servers via pip install -e, symlinks all stage skills into ~/.claude/skills/, and writes the slash commands. Restart your editor.

Claude Desktop config

{
  "mcpServers": {
    "ai-architect": { "command": "ai_architect_mcp" },
    "codebase-intelligence": {
      "command": "python3",
      "args": ["-m", "ai_codebase_intelligence"]
    }
  }
}

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

Usage

Slash commands

Direct tool calls

The MCP tools are discoverable from any Claude Code session. No slash command needed for one-off use:

Use ai_architect_codebase_query to find the OrchestrationEngine class
Use ai_architect_codebase_dead_code with min_confidence 0.9
Use ai_architect_run_hor_rules on this PRD
Use ai_architect_compound_score with relevance 0.8 uniqueness 0.6 impact 0.7 confidence 0.85

Index a codebase

ai_architect_codebase_analyze --repo-path /path/to/your/repo

This runs the 6-phase ingestion (scan → structure → parse → resolve → community → process) and persists everything to ~/.codebase-intelligence/registry.json. Subsequent queries are instant.

Architecture principles

1. Skills define WHAT. Tools define HOW. Claude decides WHY. Never mix layers.

2. The model generates. The system verifies. No LLM judges LLM output. Stage 7 is fully deterministic.

3. Adapters implement Ports. No hardcoded infrastructure calls in stage logic.

4. Context flows forward, never backward. Each stage reads upstream, writes to its own slot.

5. Every algorithm traces to a paper. No invented constants. No "industry common sense."

6. Full engine on every run. No license gates, no tier checks, no feature flags.

Tests

cd ai-codebase-intelligence
PYTHONPATH=src python3 -m pytest tests/unit/ -q
# 544 passed in ~6s

The CI workflow at .github/workflows/test.yml runs this on every PR plus a 300-line file-size guard. No PR merges without both checks green.

Project structure

ai-architect-mcp/
├── README.md
├── CLAUDE.md                     # Project instructions for Claude
├── .github/workflows/test.yml    # CI: 544 tests + file-size guard
├── .mcp.json                     # MCP server auto-discovery
├── commands/                     # Slash commands (4)
├── skills/                       # Stage skills (13 — orchestrator + 12 stages)
│   ├── orchestrator/SKILL.md
│   ├── stage-0-health/SKILL.md
│   ├── stage-1-discovery/SKILL.md
│   └── …
├── mcp/                          # ai-architect MCP server (49 tools)
│   ├── ai_architect_mcp/
│   │   ├── _tools/               # Tool definitions
│   │   ├── _verification/        # 64 HOR rules + 5 verification algorithms
│   │   ├── _prompting/           # 5 prompting algorithms
│   │   ├── _interview/           # 10 dimension scorers
│   │   ├── _adapters/            # Git, fs, xcode ports
│   │   └── server.py
│   └── pyproject.toml
├── ai-codebase-intelligence/     # codebase-intelligence MCP server (17 tools)
│   ├── src/ai_codebase_intelligence/
│   │   ├── _models/              # Typed Pydantic graph models
│   │   ├── _analysis/            # Community + process detection
│   │   ├── _extraction/          # Call routing, heritage, type extractors
│   │   ├── _resolution/          # Per-language import resolvers
│   │   ├── _search/              # BM25 + RRF hybrid search
│   │   ├── _wiki/                # LLM-driven wiki generator
│   │   └── core/
│   │       ├── ingestion/        # 6-phase pipeline (incl. ownership, cochange)
│   │       ├── analysis/         # Bus factor, churn, dead code
│   │       └── storage/          # SQLite + in-memory graph index
│   └── tests/unit/               # 544 tests
└── scripts/setup.sh              # One-command install

System requirements

• Python 3.12+

• Git

• Claude Code, Cursor, VS Code, or Claude Desktop with MCP support

Part of a bigger system

• PRD generation is delegated to ai-prd-generator — same author, same methodology, free.

• Cognitive memory is delegated to Cortex — persistent learning across sessions.

• Specialized agents are available via zetetic-team-subagents — orchestrator, engineer, code-reviewer, test-engineer, etc.

License

MIT — see LICENSE.

Troubleshooting

Commands not found — Run ./scripts/setup.sh again. Check that ~/.claude/skills/orchestrator/ exists and is a symlink.

MCP server not connecting — Verify Python 3.12+ is installed. Run python3 -m ai_architect_mcp.server to test directly. For codebase intelligence: python3 -m ai_codebase_intelligence.

Tests failing locally — Run cd ai-codebase-intelligence && PYTHONPATH=src python3 -m pytest tests/unit/ -x -v for verbose output.

MCP daemon caches stale state after analyze — Restart Claude Code after the first index of a new repo. The daemon re-reads the registry on startup.

Built by Clement Deust — methodology over tooling. The pipeline is the vehicle, the verification is the value.