Cline Coding Agent with OpenVINO Model Server#

Intro#

Cline is an autonomous coding agent extension for Visual Studio Code. It can read and edit files, run terminal commands, browse the web and call tools exposed by MCP servers, all driven by an LLM reachable through an OpenAI-compatible API. This demo shows how to serve that LLM locally with OpenVINO Model Server (OVMS) so that Cline runs entirely on your own machine, without sending code or prompts to an external service.

Requirements#

Windows (standalone package) or Linux (Docker)
Visual Studio Code with the Cline extension installed
Make sure that Cline works with next version (to change it go to Visual Studio Code settings and search @ext:saoudrizwan.claude-dev, check next option from dropdown)
Hardware: Tested on Intel Core Ultra iGPU with 32GB RAM and a discrete Intel Arc B70 GPU (dedicated VRAM). Qwen3-Coder-Next can be deployed on iGPU with 64GB RAM on board.
Memory requirements depend on the chosen model (see table below)

extension_webpage

Suggested models#

Model	HF Link	Notes
`OpenVINO/Qwen3.8-27B-int8-ov`	link	Vision-capable (VLM); general purpose chat/agent model
`OpenVINO/Qwen3.6-35B-A3B-int4-ov`	link	Vision-capable (VLM); general purpose chat/agent model
`OpenVINO/Muse-Glimmer-30B-int4-ov`	link	Vision-capable (VLM); general purpose chat/agent model
`OpenVINO/gpt-oss-20b-int4-ov`	link	General purpose chat/agent model
`OpenVINO/Qwen3-Coder-Next`	To be published soon	Big coding model; available only on iGPU with minimum 64GB of RAM
`OpenVINO/Qwen3.5-9B-int4-ov`	link	Smaller model, use when RAM/VRAM is limited or for quick, low-latency edits; not recommended for harder coding tasks

Deploy OVMS#

Windows: deploying on bare metal#

All models can be deployed with the same command, only --source_model parameter should be changed to desired model.

mkdir c:\models
ovms --model_repository_path c:\models --source_model OpenVINO/Qwen3.8-27B-int8-ov --rest_port 8000

Linux: via Docker#

All models can be deployed with the same command, only --source_model parameter should be changed to desired model.

mkdir -p ${HOME}/models
export GPU_ARGS=$(if ls /dev/dri/render* >/dev/null 2>&1; then echo "--device /dev/dri --group-add $(stat -c '%g' /dev/dri/render* | head -n1)"; fi) 
docker run -d -p 8000:8000 --rm --user $(id -u):$(id -g) -v ${HOME}/models:/models/:rw ${GPU_ARGS} \
    openvino/model_server:latest-gpu \
    --model_repository_path /models --source_model OpenVINO/Qwen3.8-27B-int8-ov --rest_port 8000

Set Up Visual Studio Code#

Install the Cline extension #

Point Cline at your OVMS instance#

Open Cline’s settings and add a new API provider configuration:

API Provider: OpenAI Compatible
Base URL: http://localhost:8000/v1
API Key: any placeholder value, e.g. unused (unless configured on OVMS server side)
Model ID: value provided as --source_model, e.g. OpenVINO/Qwen3.8-27B-int8-ov or as the --model_name e.g. ovms-model

examplary_configuration

Cline lets Plan mode and Act mode use different models/providers, so you can, for example, use Qwen3.6-35B-A3B-int4-ov for Plan-mode reasoning and switch to Qwen3.8-27B (on B70) for Act-mode code changes. You can set that checking Use different models for Plan and Act modes.

Running two large models like these at the same time can exceed available RAM/VRAM, so configure OVMS’s idle servable management (preview feature) to keep only the model used by the active mode loaded; switching modes in Cline then wakes up the other model on demand instead of requiring both to be resident at once.

Usage examples#

Chat (Plan mode)#

Ask a question about the codebase without letting Cline make changes:

Explain how request validation works in this repository and list the files involved.

Coding / agentic (Act mode)#

Switch to Act mode and give Cline a task that requires editing files and running commands:

Add a unit test for the `parseConfig` function and run the test suite to confirm it passes.

Cline will use its built-in tools (file read/write, terminal, browser) to make the change and verify it, relying on OVMS’s auto-detected tool-call format (derived from the model’s chat template) to turn the model’s tool calls into these actions.

Image input#

With Qwen3.8-27B-int8-ov, Qwen3.6-35B-A3B-int4-ov or Muse-Glimmer-30B-int4-ov deployed as a VLM, attach an image to a Cline chat message, e.g.:

[attach a screenshot of a UI bug]
What is wrong with this layout and which CSS file should I fix?

Image attachments are only available when the configured model declares supportsImages in Cline’s model configuration; enable it for the custom OpenAI-compatible model entry when using a vision-capable model like Qwen3.8-27B-int8-ov.

`reasoning_effort` usage#

Cline exposes a reasoning-effort control that is sent as part of the chat completion request. OVMS supports the OpenAI API reasoning_effort field natively, so this setting is honored across all of the suggested models above, regardless of the underlying chat template. It may be changed in Cline’s Settings under Reasoning Effort section.

Coding task#

For coding task example check CodingAgenticWorflow Demo.

Agentic and coding capabilities summary#

Agentic: Act mode combines the model’s tool-calling ability (auto-detected by OVMS from the model’s chat template) with Cline’s built-in tools (terminal, file read/write, browser) and any MCP servers you register, letting Cline plan and execute multi-step tasks autonomously.
Coding: Use Qwen3.8-27B on B70 or Qwen3-Coder-Next on iGPU for the strongest code generation/editing quality among the suggested models; fall back to Qwen3.5-9B for lightweight or latency-sensitive edits.
MCP: any MCP server (weather, filesystem, browser, etc.) becomes available to Cline the same way it is available to the OpenAI Agents SDK example in the agentic AI demo.
Image input: use Qwen3.8-27B, Qwen3.6-35B-A3B-int4-ov or Muse-Glimmer-30B-int4-ov for prompts that include screenshots or diagrams.