Skip to content

COMPATIBILITY_MATRIX.md

Latest commit

 

History

History
144 lines (112 loc) · 5.32 KB

COMPATIBILITY_MATRIX.md

File metadata and controls

144 lines (112 loc) · 5.32 KB
Raw

AMC Compatibility Matrix

A practical compatibility guide for teams deciding whether AMC will work with their stack.

What this document covers

This matrix focuses on real adoption questions:

  • Can AMC wrap my agent without code changes?
  • Can AMC capture evidence from my framework?
  • Can AMC run scoring, assurance, trace inspection, and reporting?
  • What level of support should I expect right now?

Summary

Category Status
CLI-based agent wrapping Strong
Node/Python framework support Strong
OpenAI-compatible endpoint scoring Strong
Full trust scoring + assurance flows Strong
Non-agent LLM app scoring (lite-score) Strong
Browser playground / no-install trial Available
Enterprise deployment options Available
Public hosted sandbox for real agent execution Not yet first-class

Framework compatibility

Framework / Runtime Built-in Adapter Zero-Code Wrap Evidence Capture Quickscore Assurance Packs Trace / Observe Notes
LangChain (Python) langchain-python ✅ ✅ ✅ ✅ ✅ Strong default path
LangChain (Node) langchain-node ✅ ✅ ✅ ✅ ✅ Strong default path
LangGraph langgraph-python ✅ ✅ ✅ ✅ ✅ Good support; deeper graph semantics can improve over time
CrewAI crewai-cli ✅ ✅ ✅ ✅ ✅ CLI-first support is solid
AutoGen autogen-cli ✅ ✅ ✅ ✅ ✅ Multi-agent semantics can deepen further
OpenAI Agents SDK openai-agents-sdk ✅ ✅ ✅ ✅ ✅ Strong path for agent workflows
LlamaIndex llamaindex-python ✅ ✅ ✅ ✅ ✅ Good support
Semantic Kernel semantic-kernel ✅ ✅ ✅ ✅ ✅ Good support
Claude Code claude-cli ✅ ✅ ✅ ✅ ✅ Useful for coding-agent evaluation
Gemini CLI gemini-cli ✅ ✅ ✅ ✅ ✅ CLI-based path
OpenClaw openclaw-cli ✅ ✅ ✅ ✅ ✅ First-class fit
OpenHands openhands-cli ⚠️ ⚠️ ✅ ✅ ✅ Experimental / manual install assumptions
Generic CLI agent generic-cli ✅ ✅ ✅ ✅ ✅ Fallback for basically anything shell-runnable
Plain chatbot / LLM app n/a n/a n/a ✅ Partial Partial Use amc lite-score + datasets/imported evals

Provider / model endpoint compatibility

Provider Type Status Notes
OpenAI-compatible endpoints ✅ Strong path for scoring, datasets, and gateway-backed evaluation
OpenAI ✅ Strong
Anthropic ✅ Strong
Gemini ✅ Strong
Local / self-hosted OpenAI-compatible endpoints ✅ Practical path via gateway and eval workflows
Mixed-provider fleets ✅ Supported through fleet, trace, and scoring workflows

Workflow compatibility

Workflow Status Notes
Quick trust score ✅ amc quickscore
Deep diagnostic ✅ Full question bank and scoring modules
Active red-teaming ✅ Assurance packs
Compliance reporting ✅ Audit binders, domain packs, framework reports
Trace inspection ✅ amc trace ...
Observability / anomaly views ✅ amc observe ...
Correction tracking ✅ correction log/report workflows
Golden datasets ✅ amc dataset ...
Import external evals ✅ Promptfoo / other result ingestion paths
Lite scoring for non-agent apps ✅ amc lite-score
Leaderboards ✅ Fleet comparison and export
AI asset inventory ✅ amc inventory scan
Communications policy checks ✅ amc comms-check

Environment compatibility

Environment Status Notes
macOS ✅ Strong
Linux ✅ Strong
Windows ⚠️ Usable, but some shell-centric flows may be smoother via WSL/container
Docker ✅ Supported
GitHub Actions ✅ Supported
Local developer workstation ✅ Excellent fit
CI/CD pipeline ✅ Good fit
Air-gapped / offline verification ✅ Bundle/cert verification model supports this

Recommended adoption paths

If you already run an agent framework

Use:

  • amc wrap <adapter> -- ...
  • amc quickscore
  • amc assurance run --all
  • amc trace inspect
  • amc observe timeline

If you only have a chatbot or LLM app

Use:

  • amc lite-score
  • amc dataset create
  • amc dataset run
  • amc eval import

If you are a security/compliance team

Use:

  • amc assurance run
  • amc compliance report
  • amc audit binder create
  • amc business report

Known friction points

  • Some frameworks are better supported through CLI wrapping than native SDK semantics.
  • Windows users may prefer Docker or WSL for the smoothest experience.
  • OpenHands is not as mature as the mainline adapter set.
  • Hosted browser sandbox for real execution is still more roadmap than polished product.

Bottom line

AMC already works best for teams that want to:

  • wrap an existing agent without rewriting it,
  • score trust maturity from execution evidence,
  • run red-team and compliance workflows,
  • and build repeatable evaluation loops over time.

If your system can run from a CLI, a Node/Python stack, or an OpenAI-compatible endpoint, AMC is already in very workable territory.