Four products. Three OSS surfaces. One data layer.
Bookbag ships as four tightly-integrated products that share a single org-scoped backend — plus three open-source integration paths so the runtime gate runs wherever your agent runs.
The four products
They share an organization, a RuntimeKey, and an event stream. You move between them from the same app switcher.
Observe
See every AI action
- Live activity feed
- AI inventory (apps, agents, MCP, datasets)
- Full causal trace per call
Guardrails
Block what shouldn't happen
- Visual rule builder
- Block / hold / redact / flag at call time
- Inline Create policy from any row
Evaluation
Score what did happen
- Taxonomy-driven QA
- Staged AI auditor (fast / standard / deep)
- SFT / DPO / ranking export
Governance
Prove what happened
- EU AI Act / NIST / ISO 42001 / SOC 2
- Evidence auto-mapped from runtime
- Signed audit bundles
Three ways to connect
Open source from day one. Pick the integration path that matches your agent's architecture.
Client SDK
@bookbag/sdk · pip install bookbagZero-dependency Python + Node. Drop into your agent code. Official wrappers for Anthropic and OpenAI. In-beta wrappers for LangGraph, CrewAI, AutoGen.
MCP Gateway
docker run bookbaghq/bookbag-gatewayMIT-licensed transparent MCP proxy. Sits between your agent and its MCP servers, enforces policy from a local offline-first cache, exports OpenTelemetry traces.
MCP Server
npx bookbag-mcp-serverPoint Claude Desktop, Cursor, or any MCP-native agent at a server that exposes Bookbag's gate + eval API as first-class MCP tools. stdio or streamable-http.
One API key. Explicit routing.
A RuntimeKey declares exactly which engines run when a request lands: Guardrails, Evaluation, or both. Manage keys, pick engines, pick target projects — all from Integrations → API Keys in the dashboard.
See how Integrations + API keys workReady to see your agent through the governance lens?
Join the teams shipping safer AI with real-time evaluation, audit trails, and continuous improvement.