We build the agent teams, then keep them running.

What we do

JustDukkan builds custom AI agent teams. For a company that wants a real internal process run by software, we design a hierarchical agentic system around that process, build it, connect it to the systems the company already runs, and then keep it working.

We do not resell off-the-shelf agents and we do not connect clients to someone else's agent subscription. Every system is designed for one company's processes, data and tools, and the instructions, tools, skills and integrations belong to that company.

The work splits into four parts:

• Agent team design. Turning a real process into agent roles: an orchestrator that plans and routes, specialist agents with one bounded job each, explicit limits on what each may read and change, and the points where a person decides.
• Build and deployment. Writing the agents, the typed tools they call and the skills that carry the company's procedures, then taking the team into production one capability at a time.
• Integrations: API and MCP. Connecting the team through the company's own APIs and through MCP servers that expose only what each agent is allowed to touch, with auth, audit logs and approval gates.
• Operate and maintain. Keeping a live team working as processes change, as models change and as the systems around it change.

How we work

Every engagement starts from one real process, mapped with the people who actually perform it. From that map we write the design: agent roles, tools, data access, escalation paths and the points where a person signs off. If it is worth building, we build it agent by agent, with evaluations and human checkpoints, and then keep it running.

There is no price list and no packages. After discovery we propose a fixed-scope phase with a clear deliverable and price, so the decision is made phase by phase. Maintenance is agreed separately once a system is live.

What we build with

Architectures follow a hub-and-spoke pattern: one orchestrator, specialist agents with bounded scope, typed tools and MCP servers on every edge, and human decision points where risk requires them. Repeatable procedures are packaged as versioned skills rather than buried in prompts, and every change is measured against an evaluation set built from real cases.

We design for the models companies already use, including Claude, ChatGPT, Gemini, Kimi, DeepSeek, Mistral, Llama, Qwen and Grok, with routing so each task uses the cheapest model that meets the quality bar. Cloud foundations run on AWS or Google Cloud, in the client's own account.

The reasoning behind these choices is written down in our Insights notes, which double as the reference material we use on engagements.

Open source

We publish reference implementations of the patterns we use. invoice-intake-mcp is a working MCP implementation of supplier invoice intake: separate read and commit servers, typed tools and typed errors, a server-side approval gate, an append-only audit log, a small orchestrator and a replayable evaluation set. It is MIT licensed, on PyPI, and listed in the official MCP Registry as com.justdukkan/invoice-intake-mcp.

Company

• Legal name: JustDukkan, LLC
• Registered: Delaware, USA · EIN 38-4388070
• Address: 1111B S Governors Ave, Suite 46654, Dover, DE 19904, USA
• Contact: [email protected] · +1 (910) 679-9657 · WhatsApp
• Working languages: English and Turkish
• Where: remote, with companies internationally
• Elsewhere: LinkedIn · GitHub