AI Control Framework
Part of Workshown — show your work. ai-control-framework (the how) · orchestra-lite (the scale) · (the enforcement) · (the grade) · (the front door) ·
接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
來自發布者
Part of Workshown — show your work. ai-control-framework (the how) · orchestra-lite (the scale) · (the enforcement) · (the grade) · (the front door) ·
Verification gates for AI-assisted code: deployment-readiness scoring, specs, checklists. npx ai-control-framework
Stop wasting time on non-deployable AI-generated code. Ship with confidence.
What CI actually verifies (this framework's own rule — no claim without a check): the 28-check validation suite, the MCP server build on Node 20/22, and the dashboard's 14 vitest tests on Node 20/22. An earlier "136/136 tests" badge here referenced a runner that no longer exists in the repo and has been retired.
Dogfood score: the framework run against its own repo scores DRS 27/100 (2026-07-18,
ai-framework/reference/bash/drs-calculate.sh) against its own ship threshold of 85. Published as-is: most DRS components measure live-session artifacts (frozen contracts, fresh evidence, integration proofs) that a framework repo does not carry — the score is honest, and low, and that is the point of the tool.
95% of generative AI pilots fail to move to production — MIT/Fortune, 2025
42% of AI projects were abandoned in 2025 (up from 17% in 2024) — S&P Global
This framework exists because AI coding tools are powerful but undisciplined. Without guardrails, they produce impressive demos that never ship.

Run one command. Get a deployability score. Know exactly what to fix.
AI coding assistants (Claude Code, Cursor, Copilot) fail in predictable ways. Instead of asserting frequencies, this repo measures them — on the author's own repos (methodology, pinned commits, raw data, 2026-07-18, one developer, N=19; observational, not causal — read the limitations before quoting):
Earlier versions of this README asserted "68% of sessions", "4.2x per feature", and a before/after results table. No dataset behind those numbers exists; they are retired, not re-derived. What's above is what could actually be measured.
This framework addresses 13 specific failure patterns; the benchmark measures the statically detectable ones.
The before/after table that used to sit here (time-to-deploy, rework rate, breaking changes) is retired — it had no dataset behind it. The measured evidence that exists lives in benchmarks/: what structural discipline demonstrably changes on this author's own repos is real CI (15% → 67%), licensing (38% → 100%), and the absence of mock data in production paths — and what it demonstrably does not automatically fix is README claim hygiene (the disciplined cohort was worse). Session-level outcome deltas remain unmeasured; if that measurement is ever built, it will appear in benchmarks/ with its methodology, not here as a table.
git clone https://github.com/sgharlow/ai-control-framework.git
cd ai-control-framework
# Mac/Linux
./install.sh your-project-path
# Windows PowerShell
./Install.ps1 your-project-path
bash ai-framework/reference/bash/initialize-project.sh
I'm using the AI Control Framework for disciplined development.
MANDATORY: Read these files:
1. CLAUDE.md - Operating instructions
2. ai-framework/templates/code.md - Session state
Run ./ai-framework/scripts/can-i-continue.sh now.
$ ./ai-framework/scripts/check-contracts.sh
✓ Contracts frozen: api/openapi.yaml, db/schema.sql
# Later...
$ ./ai-framework/scripts/check-contracts.sh
✗ CONTRACT VIOLATION DETECTED!
STOP: api/openapi.yaml hash mismatch
# Minute 15
$ ./ai-framework/scripts/detect-mocks.sh
⚠ 2 mocks detected - 15 minutes remaining
# Minute 35
$ ./ai-framework/scripts/detect-mocks.sh
✗ VIOLATION: Mocks detected after 30-minute mark!
Required: Replace with real service calls
$ ./ai-framework/reference/bash/drs-calculate.sh
═══════════════════════════════
DEPLOYABILITY SCORE: 85/100
═══════════════════════════════
✓ Contract Integrity (8/8)
✓ Behavioral Contracts (8/8)
✓ Security Validation (18/18)
✓ Data Integrity (10/10)
✓ No Mocks (8/8)
✓ Tests Passing (7/7)
✓ Integration Evidence (10/10)
✓ Architecture Stability (7/7)
⚠ Production Readiness (9/15)
✓ Context Preservation (8/8)
✅ READY TO DEPLOY - DRS ≥ 85
your-project/
├── CLAUDE.md # AI agent instructions
├── ai-framework/
│ ├── specs/ # 12 specifications
│ ├── reference/
│ │ ├── bash/ # Bash scripts
│ │ └── powershell/ # Windows scripts
│ ├── templates/ # 9 tracking templates
│ └── prompts.md # 20 ready-to-use prompts
└── ai-framework-mcp-server/ # MCP integration for Claude Code
# Adjust limits in check-scope.sh
MAX_FILES=5 # Files per session
MAX_LINES=200 # Lines per session
# Adjust timeout in detect-mocks.sh
MOCK_TIMEOUT=30 # Minutes
# .github/workflows/ai-control.yml
name: AI Control Framework
on: [push, pull_request]
jobs:
validate:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v2
- name: Check Contracts
run: ./ai-framework/scripts/check-contracts.sh
- name: Check DRS
run: |
DRS=$(./ai-framework/scripts/drs-calculate.sh | grep SCORE | cut -d: -f2 | cut -d/ -f1)
if [ $DRS -lt 70 ]; then exit 1; fi
Q: What if I need more than 5 files? Deploy what you have (DRS 85+), then start a fresh session.
Q: Can I override the mock timeout? No. This is intentional. 30 minutes forces real implementation.
Q: What if contracts must change?
Run approve-contract-change.sh with justification. Creates audit trail.
Q: Does this work with all AI assistants? Yes, if they can read project files. Optimized for Claude Code.
We're exploring a hosted dashboard for teams. Planned features:
Interested? Join the discussion and tell us what features matter most to you.
The core framework will always be free and open source.
See CONTRIBUTING.md.
Priority areas: language-specific patterns, CI/CD integrations, success stories.
MIT License - Use freely in personal and commercial projects.
Without this framework: 70% of AI coding time wasted on non-deployable output.
With this framework: Ship production code in hours, not days.
Stop hoping AI code will work. Start knowing it will deploy.
Verification gates for AI-assisted code: deployment-readiness scoring, specs, checklists. npx ai-control-framework
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Ai Control Framework 透過 stdio 在本機執行。
上次檢查時間:2026-10-28。 Verification gates for AI-assisted code: deployment-readiness scoring, specs, checklists. npx ai-control-framework