Workshops: Agentic Coding Tools in Depth

Local LLM Coding Agents On-Premise: Workshop for Dev Teams

When source code must not leave your network, coding agents can run on open-weight models in your own data centre or on developer workstations. This workshop shows platform and development teams how to serve coding models with Ollama or vLLM, connect agents to them and judge honestly where local models keep up with cloud models and where they fall short.

Who is this workshop for?

  • Platform, infrastructure and MLOps teams who will operate the model servers
  • Development teams in defence, critical infrastructure, finance or manufacturing with strict code confidentiality
  • Security architects who approve where code and prompts may be processed
  • Teams working in air-gapped or heavily restricted networks
  • Engineering leads who need a realistic quality and cost comparison before investing in GPUs

What you will take away

  • Serve coding models on a workstation with Ollama and for a whole team with vLLM behind an OpenAI-compatible API.
  • Select suitable open-weight models such as the Qwen Coder family or Mistral's Devstral and Codestral, and check their licences for your use.
  • Connect agents like Cline, Continue, Aider or OpenHands to your local endpoint and tune context length and tool calling.
  • Measure quality, speed and hardware needs on your own test tasks instead of public leaderboards.
  • Design a hybrid setup that keeps sensitive repositories local and routes approved work to cloud models where they are clearly better.

Agenda

  1. Module 1

    Requirements and architecture

    • Which code and data really have to stay inside?
    • Workstation, team server or GPU cluster: architecture options
    • Air-gapped operation: models, packages and updates without internet
  2. Module 2

    Serving models

    • Ollama for individual developers, vLLM for many concurrent users
    • Quantisation, context window and memory: the core trade-offs
    • Exercise: deploy a coding model and expose an OpenAI-compatible endpoint
  3. Module 3

    Choosing open-weight coding models

    • Coding models from Qwen, Mistral and other open-weight families
    • Licence terms: commercial use, self-hosting, attribution
    • Small models for autocomplete, larger ones for agentic tasks
  4. Module 4

    Agents on local models

    • Configuring Cline, Continue, Aider and OpenHands for local endpoints
    • Tool calling and long agent loops: where smaller models stumble
    • Prompting and task slicing that compensate for weaker models
  5. Module 5

    Measure, then decide

    • Building an internal test set from your own tickets
    • Comparison with a cloud model on identical tasks
    • Operating model: monitoring, model updates, capacity planning

An honest look at quality

Open-weight coding models have become remarkably capable. On long, multi-step agent tasks across large codebases, though, leading cloud models such as Claude usually still have the edge. Local models often shine at autocomplete, focused edits and clearly specified tasks.

We do not promise parity. We help you measure the gap on your own code, so you can decide which work runs locally, where a cloud model is worth it and whether more hardware would actually change the outcome.

Middle paths between local and public cloud

Fully local is not the only option for confidential code. Claude models are also available through your own tenancy on Amazon Bedrock or Google Vertex AI, and Claude Code supports those platforms. We compare the routes on control, effort and quality, and you take the assessment to your information security and data protection teams.

  • Fully local: maximum control, highest operating effort
  • Own cloud tenancy: contractual control, frontier model quality
  • EU-based provider with EU hosting: managed service with data residency

Frequently asked questions

What hardware do we need?

That depends on model size, quantisation, context length and how many developers work at once. A well-equipped workstation is often enough to experiment, while team operation usually needs dedicated GPU servers. We size it with you based on measured load, not rules of thumb.

Can Claude Code run on local models?

Claude Code is designed for Claude models. If code has to stay fully inside, we use open-source agents with local endpoints. If your own cloud tenancy is acceptable, Claude Code via Amazon Bedrock or Google Vertex AI is a strong middle path.

Can the workshop take place in an air-gapped network?

Yes, on-site. We agree in advance how models, containers and packages enter your environment through your approved import process, so participants work directly on the future target setup.

Workshop

Get in touch

Want your coding agents on your own servers, with source code that never leaves the building?

Tell us roughly how many people will join and pick dates that work – we will get back to you within 24 hours with a proposal for content, format and schedule.

Prefer to talk first?

15 minutes, free and without obligation – with our AI Trainers and AI Code Engineers.

Book a call for free (opens in new tab) +49 761 519 751 50 (24/7)

Request a training / workshop

Workshop: Local LLM Coding Agents On-Premise: Workshop for Dev Teams · 1 day

Dates that work for you (optional)

Suggestions about 2 weeks ahead – pick as many as you like

Reply usually within 24 hours · jocapps is ISO 27001 certified · Privacy