// /compare · 9 axes · 5 products
node-native, edge-ready, MIT-licensed.
How BonkLM sits next to Guardrails AI, Lakera Guard, NVIDIA NeMo Guardrails, and OpenAI’s Moderation API. Cyan ✓ ours; muted dash competitive; red ✕ absent. Last verified 2026-05-24 · methodology in /docs/compare.
AXIS
OURS
BonkLM
OSS · PYTHON
Guardrails AI
SAAS · CLOSED
Lakera Guard
OSS · COLANG
NVIDIA NeMo Guardrails
SAAS · CLOSED
OpenAI Moderation API
// BONKLM EDGE
License
MIT
Apache-2.0
Closed (SaaS)
Apache-2.0
Closed (SaaS)
All five are licensable. MIT is the most permissive.
Runtime
Node.js · edge
Python · server
SaaS API only
Python · server
SaaS API only
Node + edge is the only stack that runs in V8 isolates.
Layer count
9 layers
6 validators
4 categories
5 rails
7 categories
Layer count isn't a flex — coverage shape matters more.
Connectors
43 integrations
4 frameworks
3 frameworks
3 frameworks
1 (OpenAI SDK)
43 integrations is the largest connector surface in the field.
Edge-compatible (V8)
Yes · wasm validators
No
Network call
No
Network call
Edge-compatible means zero network hops to validate.
Self-hostable
Yes
Yes
No
Yes
No
OSS bar is met by four of five.
On-prem / airgapped
Yes
Yes
Private cloud
Yes
No
Airgapped requires self-host — SaaS-only is out.
Latency (p50)*
~0.05 ms · in-proc
Local · Python proc
Network call
Local · Python proc
Network call
Full engine, short prompt — see benchmarks. Network calls are 1000× slower minimum.
Price
Free + Enterprise
Free
Paid SaaS*
Free
Paid SaaS*
Free MIT covers most use cases — Enterprise sells access to humans, not features.
// where each one shines
Four products, honestly framed.
Feature parity isn’t the whole story. Below is the same matrix lens that ships in the upstream README: for each tool, what surface does it actually cover, what mechanism does it use, and what does it genuinely do well today — plus an honest caveat for each, including BonkLM’s own.
BonkLM
OURS · NODE · MIT- Surface coverage
- 7-surface canonical taxonomy: text_input, text_output, tool_call, retrieved_doc, memory_write, composed_context (v0.4.0); audio_partial in v0.6.
- Approach
- Deterministic pattern + structural defence. No ML in the hot path — every verdict is reproducible and runs in-process in microseconds.
- Strengths
- TypeScript-native, zero-deps core, framework / provider / platform agnostic
- Dedicated tool_call walker + handoff inputFilter + sealed wrapMemory for web3 agents
- ~0.05 ms p50 in-process (full engine, short prompt — single-core M-series, see `packages/core/benchmarks/RESULTS.md`)
- Honest caveats
- Pattern engine, not ML — multilingual coverage is regex breadth, not depth
- Stream partial-leak prevention requires full-response mode in current shape
Lakera
SAAS · CLOSED- Surface coverage
- Categorical scanners (prompt injection, harmful content, PII, etc.) via cloud API. Exact category count varies by tier — consult Lakera’s own docs.
- Approach
- Trained ML models, cloud-managed.
- Strengths
- Stronger multilingual recall via ML
- No infrastructure to run — managed service
- Honest caveats
- Network round-trip per call adds latency + new failure mode
- Vendor lock-in, per-request pricing, closed-source
LLM Guard
OSS · PYTHON- Surface coverage
- Broad scanner catalogue in a Python ecosystem; primarily input / output (exact count varies — consult upstream docs).
- Approach
- Hybrid ML + rules, Python-first.
- Strengths
- Broad scanner catalogue out of the box
- Apache-2.0, self-hostable
- Honest caveats
- Not Node.js or TypeScript-native — wrong runtime for many LLM stacks
- Fewer surface-specific factories (no dedicated tool_call / memory_write walker)
NeMo Guardrails
OSS · COLANG- Surface coverage
- Colang DSL for programmable conversation flow + rails.
- Approach
- Domain-specific DSL on top of an agentic engine.
- Strengths
- Excellent for conversational policy + rails — NVIDIA-backed
- Apache-2.0, self-hostable
- Honest caveats
- Colang learning curve; less deterministic than regex / structural rules
- Python-first — same runtime mismatch as LLM Guard for Node.js stacks
BonkLM is the deterministic Node.js-native pick for apps that need fast, predictable, composable guardrails wired into a specific connector / framework. It complements ML-based services — layer both: BonkLM for short-circuit, ML for what regex doesn’t catch.