Cadenya

Cadenya is a hosted agent runtime that layers tools, agents, and objectives on top of the APIs you already run. Rather than a framework you bolt into your application stack, Cadenya runs the agentic loop for you, so you can build, test, and improve agents without rebuilding your stack. You connect your MCP servers, OpenAPI specs, and existing endpoints through a single tool layer agents can use as they are. From there you define an agent, shape its abilities, and run objectives. It is built for teams that already have working systems and want to add agentic capability on top of them. Teams that want agents in their product usually face an awkward choice: adopt a framework and integrate and maintain it inside their own stack, or build the agentic loop themselves — handling context windows, approvals, event delivery, and model comparison along the way. Cadenya starts from the opposite assumption, that the APIs already exist and already work. Its stated purpose is to layer tools, agents, and objectives on top of the APIs you already run so you can build, test, and improve agents without rebuilding your stack. Because the runtime is hosted and model-agnostic, teams can adopt frontier models fast, test behaviors, compare approaches, and add functionality rather than complexity. You start with your stack. Cadenya connects your MCP servers, OpenAPI specs, and existing endpoints through a single tool layer agents can use, and the documentation states plainly that you do not rewrite your APIs to use it. Tools are connected using specs you already know, so existing endpoints become usable capabilities for agents as they are. Inference is kept separate from that tool layer: Cadenya is model-agnostic, so you point it at OpenRouter or any OpenAI-compatible endpoint and it uses that for inference. Each new account comes with $5 in credits on OpenRouter pre-configured for you, and after that you provide your own LLM provider credentials. That separation is what allows you to swap models, evolve behaviors, and expand capabilities while retaining infrastructure. Defining an agent means assembling concrete building blocks. An agent has assignments — individual tools, tool sets, and sub-agents — shown in the product's interface as items like a reroute shipment tool, an update ETA tool, a dispatch API tool set, and a customs broker sub-agent. It has memory layers, such as a carrier playbook or SLA policies, and a system prompt that describes the agent's role and how it should behave. Agents can dispatch sub-agents, and the model configuration for sub-agents can be changed to best suit the job, which the product describes as the most efficient approach for token usage and outcome. Objectives are then run against that configuration, and each objective keeps its trail. Token usage is managed inside the runtime rather than left to chance. Cadenya provides live token metering so you can stay on top of costs, reduce waste through progressive discovery, and improve efficiency as agents adapt. Progressive tool discovery keeps tool schemas out of the context window until the agent asks for them — only names ride along, so every request gets smaller. It is configurable: you can enable progressive tool discovery, set the maximum number of tools per search, provide search hints such as delays, reroutes, or customs, and set a rerank threshold, which can be left blank to skip reranking. Context compaction is handled out of the box, and context window compaction is also emitted as a webhook event type. Real-time behavior is a first-class part of the runtime. Webhooks and SSE push agent events into your apps as they happen, so downstream services react immediately, and the documentation notes that Cadenya makes it easy to wire agent events into your applications. Every event in your agentic loop is sent to a webhook endpoint you provide; the interface lists event types including assistant message, tool result, tool approval requested, sub-agent spawned, context window compacted, and timed out, each with a delivery status such as HTTP 204 and a completed state. When a tool call needs sign-off, approval-gated tools pause the agent and deliver a tool_approval_requested event so a person or system can approve before anything runs. Cadenya also ships Widgets that can be dropped into any frontend to enable agentic features like conversations and more, alongside SDKs in four languages. Observation and experimentation are built in. Cadenya lets you monitor outcomes and understand how behaviors take shape in the real world, on the premise that clear visibility means your agents show their worth. You can run variations — the interface shows a Default and a Canary side by side with different models, creation dates, assignments, memory layers, and system prompts — which lets you test behaviors and compare approaches without uprooting what already works. Feedback is captured against variations and objectives with sentiment-style scores, so a reroute that happened before an SLA breach scores positively while a case where the agent held at a facility when a reroute was available scores negatively. Every objective keeps its trail: tool calls, webhook deliveries, token usage, and the feedback people leave on the outcome. The benefits follow from that structure. You iterate without uprooting: swap models, evolve behaviors, and expand capabilities while retaining infrastructure. You evolve with what's next by adopting frontier models fast, testing behaviors, and comparing approaches, so the unified runtime lets you add functionality, not complexity. You experiment safely because variations and feedback let you evaluate behavior before and while real objectives run. Costs stay visible through live token metering and progressive discovery. And because agents can talk to your systems in real time and every objective keeps its trail, you can answer the question of what an agent actually did. Concrete workflows run through the product's own material. A freight shipment-exceptions agent watches for stalled deliveries and reroutes them via a dispatch API; the interface describes a system prompt for the shipment-exceptions agent that instructs it to reroute when a delivery stalls. A related feedback comment credits the agent with catching a customs hold and updating the ETA proactively, and another with escalating a frozen-goods lane correctly. Approval-gated tool calls pause for sign-off before running. Webhook deliveries notify an application endpoint as events occur. Widgets embed agentic conversation into a frontend. And canary variations let the same objectives be run against different models for comparison. Cadenya is aimed at developers and teams that already run APIs and want agentic capabilities without a rewrite. Getting started is deliberately short — kick off an agent using APIs you already have, and the first step is signing up. New accounts receive $5 in OpenRouter credits pre-configured, after which you bring your own LLM provider credentials; the team also offers a free month for those who email support@cadenya.com. API documentation is published, SDKs come in four languages, and the product runs as a hosted runtime rather than something you install and maintain in your own stack. Cadenya's value proposition is straightforward: bring agentic possibilities to life on top of the stack you already have. By layering tools, agents, and objectives over your MCP servers, OpenAPI specs, and existing endpoints, keeping inference model-agnostic, and shipping the operational pieces — context compaction, tool approvals, webhooks and SSE, widgets, SDKs, token metering, and observability — it lets teams start with one agent and grow from there.