To run Cline on a local LLM, start Ollama or LM Studio, open Cline's settings and choose the matching API provider; for llama.cpp or vLLM, choose OpenAI Compatible and enter the server's /v1 address. Give the model at least 32K tokens of context, 64K if memory allows, and pick one trained for tool calling. Roo Code shut down in May 2026; Kilo Code and Zoo Code connect similarly.
Cline is an open-source coding agent that lives in the editor: it reads files, proposes edits and runs commands, asking for approval as it goes. What has changed is the family around it. Version 4 moved Cline onto a shared SDK, Roo Code closed, and Kilo Code was rebuilt on a different core, so most set-up advice written before mid-2026 no longer matches the settings you see.
Everything below was read on 14 November 2026 from the projects’ own documentation, changelogs and source code, against Cline 4.1.26, Zoo Code 3.96.0 and Kilo Code 7.8.27. Model choice gets one section here and a full treatment in our comparison of open-weight coding models. If you prefer a terminal agent, the same exercise exists for OpenCode and for Claude Code pointed at a local endpoint.
Cline, Roo Code and Kilo Code: where each project stands
Roo Code began as a fork of Cline, and Kilo Code is usually grouped with the two. Their paths have since split.
| Project | State on 14 November 2026 | Latest release | Licence | Local model routes in its docs |
|---|---|---|---|---|
| Cline | Active. VS Code and JetBrains extensions, a CLI and a desktop app | Extension 4.1.26, 26 October 2026 | Apache 2.0 | Ollama, LM Studio, Atomic Chat, OpenAI Compatible |
| Roo Code | Shut down on 15 May 2026; repository archived | 3.54.0, 15 May 2026 | Apache 2.0 | None maintained |
| Zoo Code | Active continuation of Roo Code by former contributors | 3.96.0, 7 November 2026 | Apache 2.0 | Ollama, LM Studio, OpenAI Compatible, LiteLLM |
| Kilo Code | Active. VS Code, JetBrains and CLI; the CLI is a fork of OpenCode | 7.8.27, 5 November 2026 | MIT | Ollama, LM Studio, Atomic Chat, custom OpenAI-compatible |
Three consequences follow. Roo Code’s Marketplace page returned a 404 when we checked, so new installs go to Zoo Code or back to Cline. Kilo Code’s settings no longer look like Cline’s: the extension writes a kilo.json file shaped like an OpenCode config. And Cline now ships in several forms, so the same local server can sit behind the extension, the CLI and the desktop app.
How Cline connects to a local model
Cline has three routes to a model on your own hardware. All are chosen in the settings panel under API Provider.
| Provider in Cline | Default address | How context is set | Notes |
|---|---|---|---|
| Ollama | http://localhost:11434 | Model Context Window field, default 32,768 | Uses Ollama’s native /api/chat, so the value is sent as num_ctx on every request |
| LM Studio | http://localhost:1234 | Read-only; reported by LM Studio | Set the context when you load the model |
| OpenAI Compatible | None; you enter it | Context Window Size under Model Configuration | For llama.cpp, vLLM and other /v1 servers |
The Ollama row matters more than it looks. Ollama’s OpenAI-compatible endpoint cannot set a context size, which is why other agents need a Modelfile or a server-wide variable. Cline’s source says it deliberately uses the native API instead, because the /v1 path would load every model with the server default regardless of what you configured.
No Cline account is needed for a local runtime. The authorisation page says no key is required for Ollama or LM Studio, only that the runtime is running.
Cline with Ollama
-
Pull a model that supports tools. Cline Desktop’s first recommendation is Qwen3.8 27B, an 18 GB download in Ollama’s library:
ollama pull qwen3.8:27b -
In Cline, open Settings, then API Configuration, and set API Provider to Ollama. The model list is read from the server; if it is empty, Cline tells you it could not reach Ollama and lets you type the name.
-
Leave the base URL alone for a local server. Tick Use custom base URL only when Ollama runs on another machine; an optional API key field then appears for authenticated instances.
-
Set Model Context Window. The placeholder reads
Default: 32768. Ollama’s Cline page asks for at least 32K tokens, and Ollama’s general guidance for agents and coding tools is at least 64,000, so use 65,536 if the model and that context fit in memory. -
Check what was actually loaded while Cline works:
ollama psThe
CONTEXTcolumn should match the number you entered, andPROCESSORshould read100% GPU. A split between GPU and CPU is the most common reason for a slow session.
Ollama truncates the prompt to num_ctx on the server and does not raise an error, so a window that is too small shows up as an agent that forgets its instructions rather than as a failure message. Under the model options, Cline also exposes a request timeout that defaults to 300,000 ms; raise it if a large model takes longer than five minutes to load and answer the first request.
For the terminal, Ollama documents a launcher that sets Cline CLI’s provider, endpoint and model in one step, and offers to install the CLI through npm if it is missing:
ollama launch cline --model qwen3.8:27b
Cline with LM Studio
-
Download a model in LM Studio and load it with the context you want. From the command line:
lms load qwen/qwen3.8-27b --context-length 65536 -
Start the server from the Developer tab, or with:
lms server start -
In Cline, set API Provider to LM Studio and pick the model. The base URL defaults to
http://localhost:1234. The API key field is only for servers where you enabled Require Authentication in LM Studio.
The Context Window box in this provider is greyed out. Its tooltip explains that the value is returned by the connected endpoint, so the only place to change it is LM Studio. That has a side effect worth knowing: if LM Studio loads a model on demand because a request arrived for it, it uses the context length saved for that model in LM Studio, which may be much smaller than you intended. Load the model yourself first.
LM Studio lists the same models under different names from Ollama. Cline’s docs map qwen3.8:27b to qwen/qwen3.8-27b, devstral-small-2:24b to mistralai/devstral-small-2-2512 and gpt-oss:20b to openai/gpt-oss-20b.
Cline with llama.cpp or another OpenAI-compatible server
There is no llama.cpp entry in the provider list. Use OpenAI Compatible, which covers llama-server, vLLM and any other server with a /v1/chat/completions route.
Start the server with an explicit port, an alias and a context size. Current builds of llama-server listen on port 9931 by default, not 8080, and enable the Jinja chat template that tool calling relies on; our llama.cpp server walkthrough covers the other flags:
llama-server -hf ggml-org/Qwen3.8-27B-GGUF:Q4_K_M \
--alias qwen3.8-27b --host 127.0.0.1 --port 8080 -c 65536
Then fill in the provider:
- Base URL:
http://127.0.0.1:8080/v1. Since Cline 4.1.25 the URL is used exactly as entered, so include/v1yourself; earlier versions added it silently. - API Key: leave it empty unless you started the server with
--api-key. - Model ID: the alias,
qwen3.8-27b. - Model Configuration: set Context Window Size to the
-cvalue and, if you wish, Max Output Tokens.
Do not skip the last step. For an unknown model Cline assumes a 128,000-token window. If the server holds 65,536, Cline will keep sending history long after the server has run out of room, and its automatic compaction will start too late. Cline 4.1.21 added a recovery for one symptom of this: when a local server cuts a long reply short because the context is full, Cline now compacts the conversation and retries once.
For a GPU server shared by several developers, the trade-offs between runtimes are in our vLLM, Ollama and llama.cpp comparison.
Roo Code local LLM setup: what to use now
Roo Code cannot be recommended for a new install. Its README carries a notice that the extension was shut down on 15 May 2026 and names two successors: Zoo Code, a fork started by the Roo Code community, and Cline.
Zoo Code is the closer match if you liked Roo’s modes. Its README says the team is made up of earlier Roo contributors, and its migration guide moves your settings in two steps: export the settings file from Roo Code, then choose Import Settings in Zoo Code. The exported file can contain API keys, so treat it as a secret.
Zoo Code’s local providers behave differently from Cline’s in two ways:
-
Ollama context comes from the model, not from a setting. Zoo reads
num_ctxfrom the Ollama model and has no context field of its own. Create a variant with the window you want, then select that name in Zoo:FROM qwen3.8:27b PARAMETER num_ctx 32768ollama create qwen38-32k -f Modelfile -
Tool calling is native only. Zoo’s docs state that native tool calling is the only supported protocol, with no XML fallback, and that a model without it cannot be used.
LM Studio in Zoo Code needs the provider set to LM Studio, the model identifier, and a base URL only if you moved the server from http://localhost:1234.
Kilo Code local LLM setup
Kilo Code’s README now describes the Kilo CLI as a fork of OpenCode, and the VS Code extension stores its providers in the same kilo.json file. In the extension, open Settings, go to Providers and add Ollama or LM Studio; no key is needed. For llama.cpp or vLLM, scroll down to Custom provider, choose the OpenAI Compatible API and enter the base URL, after which Kilo fetches the model list from the server.
The file form is useful when you want the limits written down. This follows the custom model example in Kilo’s Ollama page:
{
"model": "ollama/qwen3-coder:30b",
"provider": {
"ollama": {
"models": {
"qwen3-coder:30b": {
"name": "Qwen3-Coder 30B (local)",
"tool_call": true,
"limit": {
"context": 32768,
"output": 8192
}
}
}
}
}
}
User-wide settings live in ~/.config/kilo/kilo.json and a project can carry its own kilo.json. Use ollama/<model> for a local server; Kilo’s docs warn that ollama-cloud/... is a hosted route that never touches localhost. For LM Studio the provider key is lmstudio, with options.baseURL set to http://localhost:1234/v1 if you need to change it, and the model key must match the identifier LM Studio serves.
Kilo’s Ollama page asks for at least 32K tokens of context and says to set the context window size in the provider settings. Its config example points at Ollama’s /v1 address, which cannot carry a context size, so set the server default as well and confirm the result with ollama ps:
OLLAMA_CONTEXT_LENGTH=65536 ollama serve
Two limits apply. Kilo’s inline autocomplete does not use your local model: its docs say completions go through the Kilo Gateway to Codestral. And Kilo is frank about expectations, noting that local models are more likely to get stuck in loops or misuse tools, and that requests time out after ten minutes by default.
Context window and the compact prompt
Three numbers have to agree: what the server loaded, what the extension believes, and what the model can do.
| Source | Guidance |
|---|---|
| Ollama defaults | 4K context below 24 GiB of VRAM, 32K from 24 to 48 GiB, 256K at 48 GiB or more |
| Ollama, for agents and coding tools | At least 64,000 tokens |
| Ollama’s Cline and Roo Code pages | At least 32K |
| Cline 4.1.26, Ollama provider | 32,768 unless you change it |
| Cline Desktop’s recommended models | 64K for most; 32K for Devstral Small 2 and Qwen3 8B |
| Kilo Code | At least 32K |
Context costs memory in addition to the weights. The VRAM calculator works out whether a given model and window fit, and our guide to models by VRAM tier starts from the card you own.
Does Cline still have a compact prompt?
Cline’s local models page still tells you to enable Use Compact Prompt under Settings, Features. The option was added for LM Studio and Ollama as a shorter system prompt for windows of 8K or less. In the 4.1.26 source, however, the setting is marked as removed, with a comment that the compact prompt is no longer supported, and the Features panel lists Auto Compact where it used to sit. If you cannot find the switch, that is the reason.
What remains is Auto Compact, which summarises the conversation when it nears the limit, plus the advice on the same docs page: keep tasks focused and start a new task when the context grows large. Auto Compact skips models whose context limit Cline does not know, which is one more reason to fill in the context field.
Which local models work with Cline
Cline hands the model a set of tools, such as read_files, search_codebase, run_commands and an editing tool, and every step of a task is a tool call. A model that writes good code but formats calls unreliably will stall. The settings panel carries its own warning that Cline uses complex prompts and less capable models may not work as expected.
Cline Desktop ships a list of models it recommends for agentic coding, each with the context it uses and an estimate of memory for model plus context:
| Model | Download | Memory needed | Context | Cline’s note |
|---|---|---|---|---|
| Qwen3.8 27B | 18 GB | ≈ 23 GB | 64K | Best all-round choice |
| Laguna XS 2.1 | 20 GB | ≈ 24 GB | 64K | Long sessions; quick for its size |
| Qwen3.6 35B A3B | 24 GB | ≈ 26 GB | 64K | Speed over depth |
| Devstral Small 2 | 15 GB | ≈ 21 GB | 32K | Exploring code, multi-file edits |
| gpt-oss 20B | 14 GB | ≈ 16 GB | 64K | 16 to 24 GB machines |
| Qwen3-Coder-Next | 52 GB | ≈ 56 GB | 64K | 64 GB or more |
| Qwen3 8B | 5.2 GB | ≈ 10 GB | 32K | 8 to 16 GB machines, simple edits |
These are Cline’s figures, not ours. Kilo’s own pick is qwen3-coder:30b, with the caveat that it sometimes fails to call tools correctly. For a wider view, see our notes on local models for tool calling.
Common failures and fixes
| Symptom | Likely cause | Fix |
|---|---|---|
| The agent forgets the task or repeats a step | Context smaller than the prompt; Ollama truncates silently | Raise Model Context Window; check ollama ps |
| Model fails to load, out of memory | Weights plus context exceed memory | Lower the context, pick a smaller build, unload other models |
| First reply takes minutes or times out | Cold load of a large model, or CPU offload | Preload the model; raise the request timeout; check the PROCESSOR column |
terminated: BodyTimeoutError after five minutes | A bug in versions before 4.1.25 | Update Cline |
| Connection errors with OpenAI Compatible | Base URL missing /v1, or wrong port | Enter the full URL; remember llama-server defaults to 9931 |
| Tool calls never happen | Model or chat template without tool support | Choose a tool-capable model; on llama-server keep Jinja on, or supply --chat-template-file |
| LM Studio error pointing at its developer logs | Context length set too low or too high in LM Studio | Adjust the context length there |
Privacy and telemetry settings
A local provider keeps inference on your machine. The extensions still have features that talk to the network.
- Cline telemetry is on by default in the extensions, the desktop app and the CLI. Clear Allow error and usage reporting under Settings, General. For the CLI, set
telemetryOptOuttotruein~/.cline/data/settings/global-settings.json. In VS Code, Cline also stops sending usage events while the editor’s own telemetry is off. - What Cline says it records: features used, errors, token counts, timings, and model and provider names. Its docs say code, file contents and prompts are excluded unless you use the Cline or ClinePass providers.
- Zoo Code collects usage and error data through PostHog, on by default, including the VS Code machine ID. Its privacy policy says this does not include code or prompts. Choose Disabled in its settings, or turn off VS Code’s telemetry, which it follows.
- Kilo Code telemetry is on by default and can be switched off with
"experimental": { "openTelemetry": false }inkilo.json. - Approvals: since version 4.0.0 Cline leaves command auto-approval off for new configurations. Keep it that way with a small local model, which is more likely to produce a wrong command.
.clineignoreis not a safeguard. Cline’s deprecations page says the file never prevented access, because the agent can still reach files through shell commands, and that it is being removed.
None of these switches is isolation. If the agent runs commands on a machine with credentials, see our guide to sandboxing agents that execute code. And if you only need chat and completions in the editor, not an autonomous agent, our VS Code local model guide covers the lighter options.
How VDF AI fits
VDF AI does not run, host, resell or integrate with Cline, Roo Code, Zoo Code or Kilo Code. A local model behind one of these extensions keeps one developer’s prompts on one machine. It does not give a team a shared answer to who hosts the model, which models and tools are approved, or where the record of an agent’s edits is kept; each of those lives in a settings panel on each laptop.
VDF Code is VDF AI’s governed coding assistant for organisations that need those decisions made centrally. Its product page describes extensions for VS Code, JetBrains IDEs, Neovim and Visual Studio, and deployment in a VPC, a data centre or a fully air-gapped site. For the model layer, VDF AI Router puts local open-weight models and approved external endpoints behind one API with allow and deny lists, and its air-gap mode restricts routing to local models.
Sources
Project repositories, documentation pages and model listings, each opened on 14 November 2026.
- Cline repository, changelog and releases
- Cline: local models, local models in Cline Desktop and OpenAI Compatible
- Cline: telemetry, Auto Compact, tools and deprecations
- Roo Code repository and shutdown notice
- Zoo Code repository, privacy policy, migration guide, Ollama, LM Studio and OpenAI Compatible
- Kilo Code repository and docs: Ollama, LM Studio and OpenAI-compatible providers
- Ollama: Cline, Cline CLI, Roo Code, context length and OpenAI compatibility
- LM Studio CLI: lms load and lms server start
- llama.cpp server README
- Ollama library: qwen3.8, qwen3-coder, devstral-small-2 and gpt-oss
- Qwen3.8-27B GGUF builds from ggml-org
Moving from one developer’s local agent to a governed set-up for a whole team? See how VDF Code runs on infrastructure you control, or talk to us.