Glassbox Framework
Wire runtime constitutional checks, claim extraction, and red-team audits into your agent so AI answers are verified before users see them.
Overview
Glassbox Framework is an MCP server for the Ship phase that constitutionally verifies AI answers with claim reasoning, ECS checks, red-team runs, and audit-ready trust cards.
What is this MCP server?
- Runtime constitutional verification pipeline for AI answers with claim-level reasoning
- ECS coherence checks plus dedicated constitution and red-team verification engines
- generate_trust_card assembles outputs with deterministic audit hashing and needs no API key
- Anthropic-backed engines for claim extraction and verification (GLASSBOX_MODEL, GLASSBOX_MAX_TOKENS, GLASSBOX_ECS_MODE)
- stdio MCP package @glassbox-framework/mcp v1.0.3 for Claude Code and compatible clients
- Published server version 1.0.3 on npm as @glassbox-framework/mcp
- Default verification model claude-sonnet-4-6 via GLASSBOX_MODEL
- Default per-engine token cap 2048 via GLASSBOX_MAX_TOKENS
Community signal: 8 GitHub stars.
What problem does it solve?
Agent answers can sound confident while skipping policy, facts, or internal rules—and you have no standard runtime pass to challenge claims before they reach users.
Who is it for?
Solo builders shipping Claude-powered features who want MCP-callable verification, red-team review, and hashed audit artifacts without building a custom safety stack.
Skip if: Teams that only need generic chat with no compliance story, or anyone unwilling to provision Anthropic credentials for the verification engines.
What do I get? / Deliverables
After you install the MCP server and set ANTHROPIC_API_KEY, your agent can run verification engines and emit trust cards with deterministic audit hashing on answers you are about to ship.
- Constitutionally verified assessments of draft AI answers with claim-level reasoning
- Red-team and ECS coherence results you can review before shipping copy or tool output
- Trust cards assembled with deterministic audit hashing via generate_trust_card
Recommended MCP Servers
Journey fit
Glassbox sits on the Ship shelf because it is built to gate unsafe or ungrounded model output before release, not to ideate features or run production uptime dashboards. Security is the right subphase: constitutional verification, coherence checks, and adversarial red-team passes are pre-ship trust controls for AI-generated content.
How it compares
Runtime AI verification MCP with constitutional and red-team engines—not a prompt-only skill or a generic logging monitor.
Common Questions / FAQ
Who is Glassbox Framework for?
It is for indie and solo builders using Claude Code or other stdio MCP clients who need claim-level verification and audits on AI-generated answers before release.
When should I use Glassbox Framework?
Use it during Ship when you are hardening agent outputs—run constitution checks, ECS coherence, and red-team passes on drafts you plan to show customers or stakeholders.
How do I add Glassbox Framework to my agent?
Add the npm MCP server @glassbox-framework/mcp to your client config, set ANTHROPIC_API_KEY for the verification engines, optionally GLASSBOX_MODEL and GLASSBOX_MAX_TOKENS, then invoke tools from your agent session.