What the agent does
- In the browser. The Code tab of the studio works on a GitHub repository you connected. You choose which repositories it can see, and its changes only ever arrive as pull requests.
- On your PC. The desktop app runs the same kind of agent on folders on your own machine: multi file edits shown as diffs you approve, commands and tests in a sandbox, your code never leaves the computer when you run a local model.
- Agent mode in chat. Beyond code, the agent can plan a task, call tools step by step and show its work as it goes.
What a run costs on LU Cloud
One typical run is a task that reads a few files, changes two or three and runs the tests, around fifteen steps. Credits per run and how many such runs a month’s credits pay for, on each plan:
| Model | Credits per run | Hosted | Hosted Pro | Hosted Max |
|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | about 5,900 | 152 runs | 398 runs | 847 runs |
| GLM 5.3 | about 28,500 | 31 runs | 82 runs | 175 runs |
| DeepSeek V4 Pro 0813 | about 36,400 | 24 runs | 64 runs | 137 runs |
| Qwen 3.8 Max | about 48,700 | 18 runs | 48 runs | 102 runs |
| Kimi K3.1 Pro | about 65,000 | 13 runs | 36 runs | 76 runs |
The runs per plan are alternatives, not a sum: one wallet pays for chat, code, images and video. The figures come from the same price table the meter charges from. A large repository, a long session or a model that thinks at length uses more; a prompt the provider has cached makes a run cheaper.
Without credits: the Flash models
Hosted includes 11 Flash models without credits; Pro and Max include 12. Unlimited prompt counts, up to 500,000 combined input and output tokens per account per UTC day in app sessions. All 40 chat models remain available on every plan. GLM 5.3 Flash uses credits on Hosted and is included in the daily allowance on Pro and Max. API keys and accounts without an active subscription always use credits.
On Hosted that means Llama 3.1 8B Turbo, Ling 3.0 flash, Gemma 4 26B, gpt-oss 120B and 7 more. They are good for smaller fixes and for reading code; for a long multi file change the larger models above are usually worth their credits.
Which plan for coding
900,000 credits: about 31 typical runs on GLM 5.3 or 13 on Kimi K3.1 Pro. Good to try the agent and for occasional tasks.
Start on Hosted2,350,000 credits: about 82 typical runs on GLM 5.3 or 36 on Kimi K3.1 Pro. For a few agent tasks every working day.
Start on Hosted Pro5,000,000 credits: about 175 typical runs on GLM 5.3 or 76 on Kimi K3.1 Pro. For daily heavy use on the largest models.
Start on Hosted MaxNot ready for a plan? A €15 credit pack works once, with no subscription, and its credits do not expire. See the packs.
The honest limits
Credit figures are estimates, not a quota, and a run that goes in circles costs credits like one that succeeds, so give the agent a clear task and stop it when it drifts. In the browser the agent works through GitHub only. When you run it on LU Cloud, your code is processed on our GPU partners; we never train on it and never sell it. If your code must not leave your machine, run the desktop app with a local model.
Questions
What does the coding agent do?
It works on a real project. In the browser it reads a GitHub repository you connected, plans the change, edits the files and hands the work back as a pull request, so nothing lands on your main branch without your review. In the desktop app it works on folders on your own machine, shows each edit as a diff you approve and runs commands and tests in a sandbox.
How many credits does one agent run use?
A typical run, a task that reads a few files, changes two or three and runs the tests, draws about 28,500 credits on GLM 5.3 and about 65,000 on Kimi K3.1 Pro. Hosted's 900,000 monthly credits cover about 31 such runs on GLM 5.3. These are estimates from the price table, not a quota: a big repository or a long session uses more, a cached prompt makes a run cheaper.
Is it free?
On your own hardware, yes: the desktop app and its agent are free and run on the local models you install. On LU Cloud, the Flash models included in your plan run the agent without credits inside the daily allowance, and every other model draws credits at its own rate.
Hosted, Pro or Max for coding?
Hosted (€19 a month) is enough to try the agent and for occasional tasks. If you run it every day on the large models, Pro (€49, 2,350,000 credits) or Max (€99, 5,000,000 credits) keeps you from running dry mid month. Pro and Max also include GLM 5.3 Flash in the daily allowance.
Which models can drive the agent?
All 40 chat models on LU Cloud can call tools, so each of them can drive the agent. The model picker marks every row, so you can see the price class before you pick.