AI agent development · MCP, tools, workflows

AI agents that take action in your business systems, with rules.

An AI agent is software that reads a request, decides which tools to use and acts: looks up the customer, drafts the reply, books the slot, files the ticket. That way of working is what people mean by agentic AI. We build AI agents for businesses in Canada and the US that do this inside your real systems, through governed tools, with human approval where it matters and a log of every action. Our own CRM, AgentlyLeads, runs this way from Claude and ChatGPT through 57 MCP tools.

MCP tools in AgentlyLeads
57
REST endpoints behind them
130+
Years on a Tier-1 telco backend
10
When you need this

Agents earn their keep where a chatbot stops being enough

Staff repeat the same multi-step tasks

Look up the account, check the history, draft the email, update the record. The steps are known, and agentic workflows can run them. Nobody has time to do them by hand for every lead or every ticket.

Your team already lives in Claude or ChatGPT

You want the assistant they use all day to reach your CRM or ERP safely, instead of copying and pasting between tabs.

A chatbot pilot answered questions but couldn’t do anything

Answering is step one. The value shows up when the agent can act on the answer inside your systems, with permission.

Legal or compliance is nervous about autonomy

Rightly so. Autonomous agents need approval gates, scoped permissions and an audit trail before anyone signs off on one touching customer data.

You want to expose your product to AI clients through MCP

A Model Context Protocol server lets any compatible assistant use your software as a tool. Done right, it’s a product feature. Done carelessly, it’s an open door.

Your phone line takes bookings and intake calls all day

A voice AI agent can answer the call, take the booking or the intake details and hand off to a person when the caller needs one, under the same rules as every other agent we build.

What we build

AI agent development, from tool design to governed operation

Tool and workflow design

We map the tasks the agent should take over, split them into tools with clear inputs and outputs for reliable tool calling, and decide which steps need a person in the loop.

MCP servers with OAuth

Model Context Protocol servers that expose your system’s tools to Claude, ChatGPT and other clients, secured with OAuth 2.1 and PKCE, with tenant isolation enforced on every call.

Human approval steps

Drafts the agent writes wait for a person before anything leaves the building. Irreversible actions such as payments, deletions and external messages get an explicit confirmation step.

Audit logs and permissions

Every tool call is logged with who asked, what was read and what changed. Permissions are scoped per user and per tenant, so the agent never sees more than the person using it.

Evaluation before launch

A test set of real scenarios the agent has to pass, with scored outputs, before it meets a customer. The evaluation runs again on every model or prompt change.

Integration with your systems

CRM, ERP, ticketing, calendars, email and internal APIs, through the REST or database layer you already have. If there isn’t one, we build the integration layer first.

Proof

An agent platform we run ourselves

We don’t sell agent patterns we haven’t operated. AgentlyLeads is CS2’s own product, and every rule below is enforced in it today.

Product, wholly owned

AgentlyLeads

CS2’s AI-native CRM. Sales teams run it from Claude and ChatGPT through 57 MCP tools behind an OAuth 2.1 server with PKCE: search contacts, read timelines, draft replies, book meetings. Nothing the agent writes goes out without a human approving it.

Read the case study

Product, wholly owned

Notermed

The AI scribe’s instructions read like a contract: use only what was said in the visit, never guess a diagnosis or a dose, leave a section out rather than fill it with assumptions. Every agent we build gets the same kind of contract.

Read what the AI must never do

Past performance, enterprise

Vodafone Kabel Deutschland

Agents are only as safe as the backends they call. Three years (10/2022 to 10/2025) on Vodafone Kabel Deutschland’s field-operations REST backend under the GDPR is the kind of system design that sits beneath our tool layers.

See past performance
How we work

One workflow first, then the next

  1. Pick the first workflow

    We choose one task with a clear start and end, frequent enough to matter and safe enough to learn on. Lead follow-up and ticket triage are typical starts.

  2. Design the tools and the rules

    Each tool gets a schema, a permission scope and a decision: can the agent run it alone, or does a person confirm? We write down what the agent must never do.

  3. Build and evaluate

    The tools, the MCP server or orchestration layer, and a scored test set of real scenarios. The agent doesn’t ship until it passes.

  4. Launch with a person in the loop

    The agent drafts and proposes; your team approves. As the log shows it getting things right, you decide which steps to let it run on its own.

Stack

What our agents are built with

  • Model Context Protocol (MCP)
  • OAuth 2.1 + PKCE
  • Claude and ChatGPT as agent clients
  • Provider-agnostic model layer (AWS Bedrock, OpenAI, Anthropic)
  • Structured tool schemas (JSON)
  • Next.js + TypeScript
  • Java 17 + Spring Boot
  • Prisma + PostgreSQL
  • AWS SES for outbound email
  • AWS ECS Fargate
  • CI/CD with OIDC federation
FAQ

Questions about AI agents for business

What is agentic AI, and is it different from an AI agent?

Agentic AI is the approach: systems that plan, call tools and act toward a goal under rules you set. An AI agent is one such system, built for one job such as lead follow-up or ticket triage. CS2 builds both, with approval steps and audit logs on every action.

What is the difference between a chatbot and an AI agent?

A chatbot answers. An agent answers and then does something: it calls a tool, updates a record, sends a message or schedules a task. That ability to act is where the business value is, and also where the risk is. Agent development is mostly about deciding which actions are allowed, under which conditions, and proving it.

What is MCP and why does it matter for my business?

The Model Context Protocol is an open standard that lets AI assistants such as Claude and ChatGPT use outside software as tools. If your system exposes an MCP server, your team can work with it from the assistant they already use, and your customers can too. AgentlyLeads exposes 57 tools this way behind OAuth 2.1.

What should an AI agent not be allowed to do?

Anything irreversible without a person confirming: sending external messages, moving money, deleting records, changing access rights. It should also never read data the current user couldn’t see themselves, and never invent facts to fill a gap. We write these limits down before building and enforce them in code, not in a prompt alone.

How do you test an agent before launch?

We build a set of real scenarios from your data, with the expected outcome for each, and score the agent against it. The set runs again whenever a model, prompt or tool changes. Once live, the audit log shows every action, so you can review decisions and widen or narrow permissions based on evidence.

Can the agent work with our existing CRM or ERP?

Yes, as long as the system has an API or a database we can reach. We wrap the operations the agent needs as tools with clear permissions. If your system has no usable interface, we build the integration layer first, which is standard backend work for us.

How do you keep agent costs predictable?

Agents can loop and call many tools per request, and multi-agent setups more so, so we cap steps and tokens per run, set a monthly budget per workflow, and route simple steps to smaller models. Usage is logged per tool and per tenant, so you see where the spend goes.

Do we need to replace the AI assistants our team already pays for?

No. With an MCP server, Claude or ChatGPT becomes the interface and your systems become the tools. Your team keeps the assistant it knows, and you keep control of what it can reach.

Pick one workflow. We’ll show you what an agent can do with it.

Bring the task your team repeats most. In 30 minutes we’ll map the tools, the approval points and what the first version should cover.