Keep control as you delegate more.
Cadra is the platform behind our AI agents. Agentic boards with agent and human lanes, workflows, a knowledge base per client, persistent memory per customer, schedule and event triggers, and connectors including MCP carry the work. Approvals, human takeover, tool permissions and run limits keep people in charge. Model routing puts every skill, role and agent on the right model, across closed, open-source and local models, with load balancing and failover between providers.
Imaginato is an official Cadra solution provider.
Give the conversation somewhere to go.
An agent can use your information and connected systems to move a request forward. Your process determines which actions it can take and which need a person.
Define success in the language of the business.
Give each agent a goal your team can recognise: an order ready for confirmation, a request resolved or handed to the right person, a procedure completed. Cadra connects that goal to the steps and information needed to pursue it.
People set the standard and review the result. The agent carries the work forward within that brief.
Trust starts with knowing who decides.
Give the agent enough authority to do its job. Keep the decisions that need judgment with your people. Make those boundaries part of how the work runs.
- ApprovalsPut an approval at any step you choose, or give people their own lane on the board.
- AccessTool permissions per agent decide which systems, tools and connector actions it may use.
- LimitsCost, time and tool-call limits on every run, and spend caps per agent.
- VisibilityEvery run is kept in the run log: its steps, the tools it used and its cost.
- SeparationKeep each client's work in its own isolated workspace.
Run every job on the right model.
Different work needs different models. Cadra sets the model per skill, per role and per agent, then routes, tests and balances across them, so cost, speed and quality are tuned job by job instead of fixed for the whole business.
A model for every skill, role and agent
Research, writing and review can each run on a different model, inside the same agent team.
Routing
Each task goes to the model that fits it, by task type, cost and response time. Quick sorting goes to a fast model; the answer that matters goes to a stronger one.
A/B testing
Run a challenger model against the current one on a share of real traffic, and compare quality, cost and speed before you switch.
Load balancing and failover
Spread traffic across AI providers, and fail over automatically when one slows down or stops responding, so the work keeps running.
Closed, open-source and local
Frontier hosted models, open models such as Llama, Qwen, DeepSeek, Kimi and GLM, and local models through Cadra desktop, side by side in one workflow.
Your keys, your caps
Bring your own provider keys, or run on Cadra's models with spend caps per agent.
Change models without rebuilding the agent.
When a better or cheaper model arrives, route work to it, test it on live traffic and promote it. The workflow, permissions and approvals around the agent stay exactly as they are.
For operations across regions, establish where the model can be used and where information may travel before choosing the setup. Local model processing is one part of that decision; connected systems and channels also handle data.
Give every agent a place in the business.
Cadra brings together agents, shared work, company knowledge, connections and controls. Imaginato uses that foundation to build around your process and maintain it as the work changes.
- Multi-agent teamsA coordinator agent hands work to specialist agents and brings the results together.
- Agentic boardsCards move lane by lane. Agent lanes do the routine, and people approve or take over in their own lanes.
- WorkflowsFixed sequences of steps for work that must run the same way every time, with an approval at any step.
- Knowledge base per clientBuilt from your documents, websites and records, so agents answer from your policy.
- Persistent memoryMemory per customer, merchant or case that persists across conversations, channels and handoffs.
- TriggersSchedule triggers in each recipient's local time, and event triggers from your systems. Work starts without anyone pressing a button.
- Tools and connectorsConnect your systems, including through MCP, with permissions per agent and per tool action.
- ChannelsOne agent on WhatsApp, web chat, Slack and SMS, with the same memory on every channel.
- Human handoff and approvalsA person takes over the same thread mid-conversation and hands it back. The agent resumes with its memory intact.
- GuardrailsCost, time and tool-call limits on every run, and spend caps per agent. A run stops the moment it hits a limit.
- Model routingA model per skill, role and agent, routing by task, A/B tests on real traffic, and load balancing with failover across providers. Closed, open-source and local through Cadra desktop.
- Run log and self-learningEvery step, tool call and cost is in the run log. Agents learn from approvals and outcomes, so the next run follows what worked.
The goal, the rules and the model strategy stay business decisions. We handle the build and operation behind them.