Cline

Cline is an open-source autonomous coding agent that lives in the VS Code sidebar. It reads your repo, edits files, runs terminal commands, and drives MCP servers — every change surfaced as a diff you approve before it touches disk. The split that defines it is Plan versus Act: Plan mode reads and reasons without writing anything, Act mode executes, and each mode carries its own model. Cline ships an OpenAI-compatible provider in its settings, so pointing it at Melious is a base URL, a key, and a model id. No fork, no patched build.

One thing to know up front: Cline's prompts are long. It ships a large system prompt and keeps whole files in context, so token counts run high compared to leaner agents. Prefix caching absorbs a lot of that across turns — see what's different below.

Setup

code --install-extension saoudrizwan.claude-dev

Plan and Act

Cline keeps a separate model per mode, and the toggle at the bottom of the chat box switches between them. That's worth using rather than ignoring: Plan mode reads a lot and writes nothing, so it rewards context length over raw capability, while Act mode is where tool-calling accuracy decides whether the task lands.

We'd start both on glm-5.3 and only split once you have a feel for the workload. If you do split, qwen3-coder-next (256k context, €0.17/€0.90 per M) is a reasonable Plan model at a fraction of the cost, with glm-5.3 or kimi-k2.7-code on Act.

Each mode stores its own model configuration, so if you set a different model in Plan mode, fill in its context window and prices there too.

A native provider is coming

We've opened a pull request against cline/cline adding Melious as a first-class provider. Once it ships you'll pick Melious straight from the API Provider dropdown, paste a key, and the model list populates itself — no base URL, and no hand-entered context windows or prices, because those come from /v1/models with your key.

Until then the OpenAI Compatible route above is the supported path, and it will keep working after the native entry lands. Fair warning on one detail we can't fix from our side: Cline renders prices with a hardcoded $. The numbers are the correct per-million rates, but the currency symbol will read dollars where your invoice reads euros.

Picking a model

Cline leans hard on tool calls — read file, write file, run command, then read the result and decide again. A model that drifts halfway through a tool chain will stall the agent, and that matters more here than raw benchmark scores.

46 of our 48 chat models handle tool calling. Filter for them at melious.ai/hub/models, or check capabilities.function_calling on GET /v1/models?include_meta=true. Three that hold up on long agent runs:

Smaller models are tempting for cost, but Cline's own settings panel warns that less capable models may not work as expected — that warning is earned. Start on one of the three above and trade down deliberately.

What's different

• EU residency for the whole session. Plan, Act, file edits, and terminal reasoning all run in Melious's EU provider network. Cline never builds a codebase index — it greps with ripgrep and reads files on demand — so nothing gets bulk-uploaded for embedding, and its checkpoint history stays in a local git shadow on your machine.
• Prefix caching does real work here. Cline resends a long system prompt and the same file contents turn after turn, which is exactly the shape caching rewards. It's transparent — there's no cache_control to set — and hits come back as usage.cached_tokens, billed at the cache-read rate. Repeating one small agent request while testing this guide, 128 of 181 input tokens came back as cache reads. See Models.
• Extra response fields don't reach the sidebar. Non-streaming completions carry environment_impact and billing_cost, but Cline's OpenAI client drops unknown fields. They're logged and aggregated in your usage dashboard.
• Per-developer keys with spend caps. Issue one key per developer with a monthly cap in the dashboard. An agent that loops on a failing test then hits a ceiling instead of your invoice.

When it breaks

• 401 Unauthorized — the key lost its sk-mel- prefix on the paste, or it's been rotated. Generate a new one at melious.ai/account/api/keys.
• Model ID stays a text field — Cline couldn't list models, which means the base URL or key is wrong. It must be https://api.melious.ai/v1, with the /v1.
• 404 model not found — a typo in the model id. GET /v1/models has the canonical spelling.
• Conversation truncates early — Context Window Size is unset or too low in Model Configuration. Cline defaults conservatively when it can't read the real value.
• Task cost shows $0.00 — the price fields are empty. Fill in the per-million rates from _meta.pricing.
• Cline stalls mid tool chain — the model stopped emitting valid tool calls. Swap to glm-5.3 and retry; if it persists on a strong model, the task is probably too large for one run.
• Rate limit errors on long sessions — Cline's context grows fast. Rate limits covers which plan lifts which cap.

Errors and retry patterns: Errors.