cascadeflow
Cascading runtime for AI agents. Optimize cost, latency, quality, and policy decisions inside the agent loop.
- 3.9k stars
- 885 forks
- 12 open issues
- Language: Python
- License: MIT
- updated
Install
git clone https://github.com/lemony-ai/cascadeflow && cp cascadeflow/*.md ~/.claude/agents/- Clone the repository and copy the agent .md definitions into ~/.claude/agents (or .claude/agents inside a project).
- Start a new Claude Code session to load the agents.
- Delegate work to them with the Task/Agent tool or by name.
Editor's note
CascadeFlow is an in-process runtime intelligence layer for AI agents that handles model selection, cost control, latency optimization, and policy enforcement inside the agent execution loop rather than at the HTTP boundary. It integrates with Claude through the Anthropic API and works alongside frameworks including LangChain, CrewAI, PydanticAI, Google ADK, n8n, Vercel AI SDK, and OpenAI Agents SDK, available as both a Python package and a TypeScript npm module. The core mechanism is model cascading: routing each agent step or tool call to the most appropriate and cost-effective model based on task complexity, token budgets, quality thresholds, and business KPIs accumulated across the run. A standout benchmark result from the README shows 93% cost reduction on GSM8K math tasks while retaining 96% of GPT-5 quality. The library adds under 5ms overhead per call and supports per-tool-call budget gating and runtime stop, continue, or escalate decisions. It targets developers and teams running multi-step Claude-based agents who need transparent cost accounting and adaptive model routing without an external proxy.
1 items in this repository
Use when building, extending, or debugging AI agents with cascadeflow (agent runtime intelligence layer) — installing `cascadeflow` (Python) or `@cascadeflow/core`/`@cascadeflow/langchain` (TypeScript); using `CascadeAgent`, `ModelConfig`, harness APIs (`cascadeflow.init`, `cascadeflow.run`, `@agent` from `cascadeflow.harness`, `simulate`), `withCascade`/`CascadeFlow`; picking drafter+verifier pairs; per-step budget/compliance/KPI enforcement; quality validation; complexity pre-routing; tool execution and multi-turn agent loops; presets; decision traces; or wiring cascadeflow into LangChain, OpenAI Agents, CrewAI, PydanticAI, Google ADK, n8n, or Vercel AI SDK. Also when a user mentions "cascade", "drafter/verifier", "runtime intelligence", "in-process harness", "cost-optimized agent", "agent loop with cost control", is in the lemony-ai/cascadeflow repo, or found a bug in cascadeflow/integrations needing an upstream fix/PR.
Subagents overview
What people ask about cascadeflow
What is lemony-ai/cascadeflow?
lemony-ai/cascadeflow is subagents for the Claude AI ecosystem. Cascading runtime for AI agents. Optimize cost, latency, quality, and policy decisions inside the agent loop. It has 3.9k GitHub stars and its last recorded update is dated 2026-09-25.
How do I install cascadeflow?
You can install cascadeflow by cloning the repository (https://github.com/lemony-ai/cascadeflow) or following the README instructions on GitHub. ClaudeWave also provides quick install blocks on this page.
Is lemony-ai/cascadeflow safe to use?
Our security agent has analyzed lemony-ai/cascadeflow and assigned a Trust Score of 97/100 (tier: Verified). See the full breakdown of passed checks and flags on this page.
Who maintains lemony-ai/cascadeflow?
lemony-ai/cascadeflow is maintained by lemony-ai. The last recorded GitHub activity is dated 2026-09-25, with 12 open issues.
Are there alternatives to cascadeflow?
Yes. On ClaudeWave you can browse similar subagents at /categories/agents, sorted by popularity or recent activity.