Glassbox

AI answers, audited by code — not vibes.

What GLASSBOX does

Glassbox is a super-thinking + auditing layer for AI coding agents. When your AI plans something, Glassbox turns its reasoning into a structured artifact — and then questions it.

It forces every plan through a recursive verification loop, scores trust 0–100 with a plain-English verdict, writes the receipts (every assumption, every risk, every step's acceptance check), and renders the whole thing as a neon HTML cockpit — so anyone can see why the AI believes its answer.

No claims without receipts. No plan without a trust score. No reasoning without a verifier signing off.

Works natively with your AI CLI

Related MCP server: agent-tasks

Install — 60 seconds, zero babysitting

Windows (PowerShell):

git clone https://github.com/DERYK-RYLOY/glassbox.git
cd glassbox
powershell -ExecutionPolicy Bypass -File .\install.ps1

macOS / Linux:

git clone https://github.com/DERYK-RYLOY/glassbox.git
cd glassbox
bash ./install.sh

The installer: checks Python, installs dependencies, finds/installs Lua, wires Glassbox into OpenCode (or prints the one-line config for your CLI), and verifies itself with glassbox_status. If something is missing, it tells you exactly what to do — it never installs anything you didn't consent to.

Manual setup (any CLI that speaks MCP):

{ "mcp": { "servers": { "glassbox": {
  "type": "local",
  "command": ["python", "<path-to-glassbox>/python/glassbox_server.py"]
} } } }

The toolkit

12 tools. One JSON contract. Works everywhere — including CLIs with no MCP at all:

# Universal CLI (no MCP, no protocol)
python python/glassbox_cli.py --list
echo '{"title":"ship","goal":"x","steps":[]}' | python python/glassbox_cli.py glassbox_trust_report

Built on

                    

The 10-second architecture

your AI CLI ──MCP──► glassbox_server ──► Iron Swarm v2 (pure Lua, recursive)
                                             3 scouts ∥ ─ Thinker gate ─► 5 analysts ∥
                                             architect ─► verifier (GV|PASS / GV|REVISE ↻)
                                                │
                                    final plan → trust_report(0–100) → receipts

• Agents speak GB1, a compact tagged dialect — not English prose. ~60–80% less wire chatter.

• The Tiny Thinker (4.2 MB, int8 BERT-Tiny) triages signal vs. noise before your model spends tokens on it.

• The recursive verifier refuses to ship until the plan survives audit — or reports exactly why it can't.

Every claim auditable · every assumption receipted · trust 98/100.

MIT licensed. Built with Lua, Python, and stubborn honesty.