Documentation
RESK documentation
Reference manuals for the RESK open-source stack. Every page is reviewed against the source code of the release it documents — including the limitations, which are part of the interface.
On this page
Manuals
ReskSafety
Self-hosted LLM firewall: OpenAI/Anthropic-compatible API, RBAC, policies, logits filtering, ENV Match masking.
resk-mcp
Security layer for Model Context Protocol servers: JWT, PII/injection screening, rate limits, dashboard.
resk-llm
Python prompt-injection pipeline (resk2), nine detectors, output validation, LLM Guard shim.
resk-llm-ts
TypeScript detectors for Node and Bun, Express/Hono middleware, OpenAI wrapper.
resk-logits
Hugging Face logits processors: banned tokens, forced phrases, length control, vLLM shim.
resksecure
Bitmask-to-policy logits filtering and tool gating. Proprietary licence.
ReskPoints
Agent action logging with field masking, sampling and console/file/webhook/Datadog/Prometheus exporters.
honeycrawlpot
Deception middleware: decoy secrets, canary tokens and prompt-injection payloads for LLM crawlers.
resk-caching
Semantic LLM response cache in Bun with memory/SQLite/Redis backends and a JWT API.
Conventions
- Source of truth. Each manual is written against the current
mainbranch and states the version it covers. Where the shipped package differs from the repository, the difference is called out. - Code samples target Python 3.11+ or Bun/Node 18+ as appropriate, and use the package's real import names (
mcp-resk→resk_mcp,resk-llm→resk2). - Limitations are documented. A control that only catches pattern-shaped attacks is described as such; a known packaging or CLI issue is listed rather than hidden.
- Versioning. The version shown at the top of a manual is the one the text was verified against.
Long-form guides
Architecture, threat models and deployment walkthroughs live outside the reference manuals: the MCP security hub, the LLM security hub, the MCP integration guide and the ReskSafety launch article.