Rifft

Rifft finds the handoff that broke your multi-agent run — not just the error message. It walks backwards through spans to the root cause, classifies the failure with the UC Berkeley MAST taxonomy, and shows you the exact agent and message where things went wrong.

Install

JavaScript SDK:

npm install @rifft-dev/rifft

Python SDK:

pip install rifft-sdk

Python adapters:

• LangGraph: rifft-langgraph
• CrewAI: rifft-crewai
• AutoGen / AG2: rifft-autogen
• MCP: rifft-mcp

Claude Code — trace every session with zero code changes:

rifft-claude init --project-id YOUR_PROJECT_ID --api-key YOUR_API_KEY

Why not Langfuse or LangSmith?

Langfuse and LangSmith are excellent general-purpose LLM observability platforms. Rifft is different in one specific way: it is built for multi-agent failure analysis, not span collection.

• Root cause, not span trees. When a 4-agent run fails, Langfuse shows you a tree of 200 spans. Rifft shows you which agent caused the failure and why, in plain English.
• MAST classification. Every failure is automatically classified against the UC Berkeley MAST taxonomy — 15 standardised failure modes covering tool call hallucinations, dropped handoffs, and context overflows. You get a label, an explanation, and a suggested fix.
• Agent graph, not a waterfall. Rifft renders the agent communication graph so you can see which agent sent what to which, where messages were dropped, and which handoff edge was the bad one.
• Not for single-LLM tracing. If you're tracing individual LLM calls in a monolithic app, use Langfuse or Helicone. Rifft is opinionated: it is for multi-agent systems where the failure could be anywhere in the chain.

What Rifft does

• Cross-framework trace ingestion over OTLP
• Agent-to-agent communication graph
• MAST failure classification with fix suggestions
• Causal attribution — root cause agent, failing agent
• Timeline and per-agent debugging views
• Dataset and eval workflow for regression testing
• Self-hosted (Docker Compose) and cloud (rifft.dev)

What Rifft does not do

• Prompt management UI
• LLM evaluation or scoring
• AI gateway or proxy features
• General APM or application monitoring
• Single-LLM call tracing as the primary use case

Framework support

Framework	Status
CrewAI	Full
AutoGen / AG2	Full
MCP	Full
Claude Code	Full (via rifft-claude)
Custom agents via SDK	Full
LangGraph	Full

5-minute CrewAI quickstart

pip install rifft-sdk rifft-crewai

import rifft
import rifft.adapters.crewai

rifft.init(project_id="my-project", endpoint="http://localhost:4318")

# Your existing crew code unchanged
crew = Crew(agents=[...], tasks=[...])
result = crew.kickoff()
# Open http://localhost:3000 to see the trace

Other Python installs:

pip install rifft-sdk rifft-autogen
pip install rifft-sdk rifft-mcp

Self-host in under 5 minutes

git clone https://github.com/rifft-dev/rifft.git
cd rifft
docker compose up -d --build
open http://localhost:3000

Default local endpoints:

• Web UI: http://localhost:3000
• API: http://localhost:4000
• Collector HTTP: http://localhost:4318
• Collector gRPC: localhost:4317

Monorepo layout

rifft/
├── apps/
│   ├── api
│   ├── site
│   └── web
├── infra/
│   └── docker
├── packages/
│   ├── adapters/
│   │   ├── autogen
│   │   ├── crewai
│   │   ├── langgraph
│   │   └── mcp
│   ├── collector
│   ├── sdk-js
│   └── sdk-python
└── docs/

Community

Ask questions, share broken multi-agent traces, and follow updates in GitHub Discussions.

License

MIT