Cadenya

Cadenya is a hosted agent runtime that layers tools, agents, and objectives on top of the APIs you already run. It is explicitly not a framework you bolt into your application stack; instead, Cadenya runs the agentic loop for you. You connect your MCP servers, OpenAPI specs, and existing endpoints through a single tool layer that agents can use, then define an agent, shape its abilities, and run objectives against it. The product is positioned for developers and teams who want to bring agentic possibilities to life in software they already operate, testing safely and improving quickly without rebuilding the stack that supports them. The problem Cadenya addresses is the cost of adopting agentic capabilities inside a system that already works. Building agents typically means invasive change: framework glue inside the application, wrapper layers around APIs, and hand-built machinery for streaming, approvals, retries, context limits, and visibility into what an agent actually did. Cadenya's answer is a hosted loop that handles those concerns out of the box. The product states that it handles context compaction, tool approvals, webhooks and SSE streaming, embeddable widgets, and SDKs in four languages. A frequently asked question on the site answers directly that you do not need to rewrite your APIs: connect MCP servers, OpenAPI specs, and existing endpoints through a single tool layer, and agents can use them as they are. That preserves existing infrastructure investment while new agentic behaviour is layered on top, which matters because teams can iterate on agent behaviour without destabilising the services their business already depends on. The first layer is the tool layer, and the product's guidance is to start with your stack. You connect MCP servers, OpenAPI specs, and existing endpoints through a single tool layer agents can use. Because connection is based on specifications developers already know, existing services become available to agents as they are. Tools can be assigned to an agent individually, organised as tool sets, and combined with sub-agents. In the interface example shown on the site, a shipment-exceptions agent carries assignments such as a reroute shipment tool, an update ETA tool, a Dispatch API tool set, and a Customs Broker sub-agent, alongside memory layers such as a Carrier Playbook covering SLA policies. This layer is what lets an agent reach into real systems rather than only producing text. Experimentation is handled through variations. The interface shows a Default variation and Canary variations, each tied to a specific model, such as Anthropic Claude Sonnet as the default and OpenAI GPT-5.5 or GPT-5.2 as canaries, with a creation timestamp. A variation also carries its own system prompt, its own assignments, and its own memory layers. The product's stated goal for this area is to let you iterate without uprooting: swap models, evolve behaviours, and expand capabilities while retaining infrastructure. It also frames the runtime as a way to evolve with what is next, so you can adopt frontier models fast, test behaviours and compare approaches, and add functionality rather than complexity. Practically, that means model and behaviour changes are configuration inside the runtime rather than a rewrite of the surrounding application. Cost and context are managed through the token usage layer. Cadenya provides live token metering so teams can stay on top of costs, and describes reducing waste through progressive discovery, with efficiency improving as agents adapt. Progressive tool discovery can be enabled, with a maximum number of tools per search set between 1 and 10 (5 shown in the example), search hints of up to 5 terms, and a rerank threshold between 0.0 and 10 that can be left blank to skip reranking. The mechanism is described precisely: tool schemas stay out of the context window until the agent asks for them, and only names ride along, so every request gets smaller. That matters for agents with many connected tools, where schema bloat would otherwise consume context and budget before the agent does any work. Cadenya is also built to power real-time services. Webhooks and SSE push live updates, and the product states it makes it easy to wire agent events into your applications. The webhook delivery view on the site lists event types including assistant message, tool_result, tool_approval_requested, sub_agent_spawned, context_window_compacted, and timed_out, each with an HTTP status and target URL. Around this, the product provides tool approvals: approval-gated tools pause the agent and deliver a tool_approval_requested event so a person or system can approve before anything runs. It also ships embeddable widgets, described as a feature called Widgets that can be dropped into any frontend to enable agentic features like conversations and, per the product, so much more. Observation closes the loop. Cadenya is designed to let you monitor outcomes and understand how behaviours take shape in the real world, with the stated goal that clear visibility means your agents show their worth. Every objective keeps its trail: tool calls, webhook deliveries, token usage, and the feedback people leave on the outcome. The feedback view shows comments scored by sentiment, attributed to a variation and an objective, such as a reroute completed before an SLA breach, a customs hold caught with a proactive ETA update, a reroute that notified the recipient twice, a hold that occurred when a reroute was available, and a correctly escalated frozen-goods lane. Those scored outcomes are what make comparison between variations meaningful. Overall the product works as a hosted agentic loop in three steps: define an agent, shape abilities, and run objectives. Inference is model-agnostic. You point Cadenya 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; after that you provide your own LLM provider credentials. Agents can dispatch sub-agents, and the model configuration for sub-agents can be changed to be best suited for the job, which the product describes as giving the most efficient token usage and outcome. The runtime itself is described as unified: you add functionality, not complexity. The benefits follow from that design. Because the loop is hosted rather than embedded, teams can experiment safely and improve quickly. Because APIs are not rewritten, capabilities expand while infrastructure is retained. Because variations, memory layers, and assignments sit in configuration, frontier models can be adopted fast and behaviours compared rather than guessed at. Because events flow through webhooks and SSE, downstream services react immediately instead of polling. Because token metering and progressive tool discovery are built in, spend is visible and requests stay smaller. And because every objective keeps a trail of tool calls, webhook deliveries, token usage, and human feedback, improvement is grounded in recorded outcomes rather than impressions. Concrete scenarios appear throughout the product's own material. The site's worked example is a freight and logistics agent named for Meridian, a shipment-exceptions agent instructed to reroute a stalled delivery via a dispatch API, with assignments covering rerouting, ETA updates, a Dispatch API tool set, and a Customs Broker sub-agent, and memory layers holding a Carrier Playbook and SLA policies. Feedback entries describe rerouting before an SLA breach, catching a customs hold and updating the ETA proactively, and escalating a frozen-goods lane. Other described workflows include dropping Widgets into a frontend for conversations, pushing agent events into applications via webhooks and SSE so downstream services react immediately, gating tool calls behind approvals, and dispatching sub-agents for parts of a job. On integrations, Cadenya connects MCP servers, OpenAPI specs, and existing endpoints through a single tool layer, uses OpenRouter or any OpenAI-compatible endpoint for inference, and delivers webhooks and SSE streams to endpoints you provide. It ships SDKs in four languages and embeddable widgets for frontends. The site notes that each new account includes $5 in credits on OpenRouter pre-configured, after which you supply your own LLM provider credentials, and states that you can email support@cadenya.com to get a free month. Getting started is described as a few steps, the first of which is signing up, with sign-up available at app.cadenya.com and API documentation linked from the site. Cadenya's core promise is that you can bring agentic possibilities to life on the stack you already run. It converts agent development from a rebuild project into a hosted runtime you configure, connect, observe, and improve, with the surrounding concerns of context, approvals, streaming, cost, and feedback handled as part of the loop.