
Glassbox Framework
Turns every AI answer into a structured Trust Card with traceable reasoning chains, epistemic confidence scores, and adversarial red team probes. Six tools exposed: verify a complete question-answer pair, extract atomic claims with their reasoning, calculate ECS breakdowns using a published formula, run seven adversarial probes (fabrication, bias injection, context attacks), assemble cards from prebuilt components, and export deterministic SHA-256 audit logs. Lets you pass deployer intents as natural language constitutional rules that get evaluated at runtime. Built as a Node MCP server with a thin Python client wrapper. Reach for this when you need runtime verification with full audit trails instead of opaque scoring, especially if you're shipping AI features where trust and traceability matter more than speed.
- GitHub stars
- 8
- tools
- 6
- auth
- Secret needed
- transport
- STDIO
- env vars
- 4
- last updated
- May 25, 2026
Tools
6 toolsPublic tool metadata for what this MCP can expose to an agent.
glassbox_verify_answerFull pipeline → Trust Card
Full pipeline → Trust Card
No parameter schema in public metadata yet.
glassbox_extract_claimsAtomic claims with reasoning chains
Atomic claims with reasoning chains
No parameter schema in public metadata yet.
glassbox_score_ecsECS with full breakdown + formula
ECS with full breakdown + formula
No parameter schema in public metadata yet.
glassbox_red_teamGlassbox Court — 7 adversarial probes
Glassbox Court — 7 adversarial probes
No parameter schema in public metadata yet.
glassbox_generate_trust_cardAssemble a Trust Card from prebuilt parts (no LLM call)
Assemble a Trust Card from prebuilt parts (no LLM call)
No parameter schema in public metadata yet.
glassbox_export_audit_reportFull pipeline + deterministic SHA-256 audit log
Full pipeline + deterministic SHA-256 audit log
No parameter schema in public metadata yet.
Sponsors
Configuration
ANTHROPIC_API_KEY*secretYour Anthropic API key (sk-ant-...). Required for the claim extractor, red team, constitution engines, and ECS coherence checks. One v1 tool (generate_trust_card) works without an API key — it's pure assembly + deterministic audit hashing.
GLASSBOX_MODELdefault: claude-sonnet-4-6Override the Claude model used by the verification engines. Defaults to claude-sonnet-4-6.
GLASSBOX_MAX_TOKENSdefault: 2048Per-engine-call max-tokens cap. Defaults to 2048; raise for verifying long-form content.
GLASSBOX_ECS_MODEdefault: arithmeticECS aggregation mode: arithmetic (weighted mean, default) or geometric (stricter — any zero dimension collapses the total).
Readme
Glass Box Framework
Runtime constitutional verification for AI answers. Every claim carries a reasoning chain. Every score breaks down. Every verdict is traceable.
Details
- Repository
- github.com/thebarmaeffect/glassbox
- Category
- Cloud & Infrastructure
- Package
@glassbox-framework/mcp- Transport
- STDIO
- Auth
- Required
- Registry
- MCP registry, active