Lemony
Also known as: cascadeflow
Lemony builds cascadeflow, an open source runtime that sits inside AI agents and scores every model and tool call on cost, latency, quality, budget, compliance and energy, then lets the step through, switches the model, denies the tool or stops the run. cascadeflow Studio adds a managed interface, a policy builder and fleet cost analytics, sold by demo.
Lemony Inc., based in New York with a presence in Zurich, builds cascadeflow, an MIT licensed runtime installed from pip or npm that runs in process inside the customer's own agents. It tries a cheaper model first, scores the draft and escalates to a larger model only when the draft fails its checks, with a domain pipeline that sends code, math and other query types to specialist model cascades.
A harness applies budgets, tool call caps, latency and energy limits and GDPR, HIPAA and PCI compliance gating, in an observe mode that only logs or an enforce mode that switches models, denies tools or stops the run, and every decision is written to a per step trace. cascadeflow works with LangChain, the OpenAI Agents SDK, CrewAI, Google ADK, PydanticAI, n8n, the Vercel AI SDK and Hermes Agent, over more than 17 model providers including local vLLM and Ollama, and it also runs as an OpenAI compatible gateway server. cascadeflow Studio is the managed version, with a UI and CLI, a Policy Builder and Domain Builder, fleet cost analytics, role based access, SSO and SAML, and audit logs, sold through a demo. Lemony previously sold an on premise AI appliance, which it no longer offers.
Vendor details
Canonical URL
https://www.lemony.ai/
Category
Agent infrastructure
Company status
independent
Use cases & customers
Integrations
LangChain and LangGraph, OpenAI Agents SDK, CrewAI, Google ADK, PydanticAI, n8n, Vercel AI SDK and Hermes Agent, over more than 17 model providers including OpenAI, Anthropic, Groq, Ollama and vLLM.
In practice
A team running LangChain agents in production wants to cut model spend without rewriting them. It adds cascadeflow in observe mode to see which steps a cheaper model could handle, then switches to enforce mode so drafts from a small model go through and only failed drafts escalate.
A healthcare company must keep certain requests away from models that do not meet HIPAA terms. It sets cascadeflow's compliance gating to hipaa, so a step headed for a disallowed model is switched or stopped, and the per step trace records why.
A platform team gives each customer facing agent a budget per run and a cap on tool calls. cascadeflow stops a run that would overspend, denies further tool calls past the cap, and reports cost, steps and tool calls for each session.
Agentic Index coverage score
8.0 / 14 capabilities · 57%
| Integrations & Tool Calling | Full |
|---|---|
|
cascadeflow runs tool calls inside the agent loop: customers define their own tools with ToolConfig and a handler, and CascadeAgent executes them across turns through a ToolExecutor. It also plugs into LangChain, the OpenAI Agents SDK, CrewAI, Google ADK, PydanticAI, n8n, the Vercel AI SDK and Hermes Agent. SourceLemony, docs.cascadeflow.ai agent loop and integrations pagesread 2026-11-17 |
|
| Workflow Orchestration | Partial |
|
cascadeflow runs multi turn agent loops capped by max_steps and lets a main agent delegate to a sub agent wrapped as a tool, with one budget across the agent tree. It documents no workflow model with branching, handoffs or retries, since it governs steps inside the customer's own framework. SourceLemony, docs.cascadeflow.ai multi agent exampleread 2026-11-17 |
|
| Knowledge Grounding & RAG | Not documented |
|
cascadeflow routes and governs model and tool calls and documents no retrieval, index or knowledge base over the customer's documents. Its domain pipeline classifies each query to pick specialist models, which is routing rather than grounding. SourceLemony, github.com/lemony-ai/cascadeflowread 2026-11-17 |
|
| Human Oversight & Guardrails | Partial |
|
In enforce mode the harness applies hard caps on budget, tool calls, latency and energy and compliance gating for GDPR, HIPAA and PCI, answering with switch_model, deny_tool or stop, and observe mode logs the decisions without applying them for a staged rollout. The cascadeflow site lists require_approval among its actions, but the documentation describes no human approval step. SourceLemony, docs.cascadeflow.ai agent harness page and cascadeflow.airead 2026-11-17 |
|
| Security, Identity & Governance | Partial |
|
cascadeflow Studio lists role based access, SSO and SAML, audit logs and multi org workspaces. Lemony names no attestation, certification or trust page, and the open source runtime's compliance gating is a model policy rather than an access model. SourceLemony, cascadeflow.ai and lemony.airead 2026-11-17 |
|
| Observability & Auditability | Full |
|
Every harness decision writes a trace record with the action, a machine readable reason, the model, the step, cumulative cost, latency, energy and budget state, and whether it was applied, read through session.trace(), session.summary() or a callback manager. The README adds OpenTelemetry export and LangSmith tracing. SourceLemony, docs.cascadeflow.ai decision trace page and GitHub READMEread 2026-11-17 |
|
| Memory & State Persistence | Partial |
|
Inside a run, CascadeAgent keeps the message and tool call list across turns, and the run session carries cumulative cost and budget state across steps and sub agents. No memory layer carries context from one run to the next. SourceLemony, docs.cascadeflow.ai agent loop page and GitHub READMEread 2026-11-17 |
|
| Deployment & Data Residency | Full |
|
The open source runtime installs from pip or npm and runs in process inside the customer's own agents, with a self hosted gateway server and local models through Ollama and vLLM. Studio is managed, and its hosting and data location are not stated. SourceLemony, github.com/lemony-ai/cascadeflow and cascadeflow.airead 2026-11-17 |
|
| Prebuilt Agents, Templates & Packs | Partial |
|
cascadeflow ships compliance presets (gdpr, hipaa, pci and strict), provider presets and about 50 Python and 47 TypeScript examples that the customer adapts. It ships no prebuilt agents. SourceLemony, docs.cascadeflow.ai examples catalog and harness pagesread 2026-11-17 |
|
| Triggers & Channel Coverage | Not documented |
|
cascadeflow runs when the customer's own code or framework calls it, in process or through its OpenAI compatible gateway server. It has no schedules, webhooks, event triggers or inbound channels of its own. SourceLemony, github.com/lemony-ai/cascadeflow and docs.cascadeflow.airead 2026-11-17 |
|
| Model Flexibility & Routing | Full |
|
The customer sets the models in each cascade from more than 17 providers, including OpenAI, Anthropic, Groq, Together, Hugging Face and on premises vLLM and Ollama, and sets KPI weights for quality, cost, latency and energy that steer which model each step uses. SourceLemony, docs.cascadeflow.ai providers and presets page and GitHub READMEread 2026-11-17 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
cascadeflow publishes a Python API and a TypeScript package, @cascadeflow/core, each with a full reference, plus an OpenAPI spec. SourceLemony, docs.cascadeflow.ai API referenceread 2026-11-17 |
|
| Testing, Debugging & Optimization | Partial |
|
Each cheap model draft is scored on length, logprob confidence, format and optional semantic alignment before cascadeflow escalates, and observe mode shows what a policy would do on live traffic before it is enforced. Studio's auto benchmarks and automatic rollback appear only on the marketing site, with no documented harness behind them. SourceLemony, GitHub README, docs.cascadeflow.ai and cascadeflow.ai; github.com/lemony-ai/cascadeflowread 2026-11-17 |
|
| Browser & Computer Use | Not documented |
|
cascadeflow governs model and tool calls and has no browser, desktop or computer control of its own. SourceLemony, github.com/lemony-ai/cascadeflowread 2026-11-17 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Free and open source under MIT; Studio by demo
What is public
Lemony's cascadeflow runtime is free. cascadeflow Studio, the managed version with a UI and CLI, policy and domain builders and fleet cost analytics, is sold after a demo, with no price on the site.
Cost watchouts
Every model call cascadeflow routes is still billed by the model provider, and cascadeflow reports that cost for each run.
Variable cost rationale
The runtime exists to lower model spend. It sends each step to the cheapest model whose draft passes its checks and enforces budget caps per run and per user.
Sales call required
Yes, required for paid access
Free / trial
The runtime is open source under the MIT license and installs from pip or npm.
Commercial notes
Studio adds role based access, SSO and SAML, audit logs and multi org workspaces to what the open source runtime does.
Missing data
Lemony does not publish a price, a billing unit or contract terms for Studio.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…
Alternatives to Lemony
Other products that do the work Lemony does, and what each one is.
- LiteLLMOpen source gateway and SDK
- LLM GatewayGateway with cost per coding agent
- PortkeyGateway with guardrails and logging
- OpenRouterMarketplace of 500+ models
- TrueFoundryEnterprise gateway, any deployment
- HeliconeObservability first gateway