Skip to content
SwisserAI LogoSwisserAI
SwisserAI for VS Code HelpSwisserAI HelpOpen app

Work

Credits and saving tips

Every step the agent takes costs a few credits. How much depends mostly on how much it has to read. These habits keep it low.

How credits are counted

A task is not one request. The agent reads, thinks, edits and checks, and every one of those steps is a request to SwisserAI. Each request costs credits by how much text goes in and out:

  • What it reads: your task, the files it opened and everything said before in this chat. This is usually the biggest part.
  • What it writes: its answer and the code of an edit.
  • Reused text is cheap: text the agent already read a few minutes ago costs a fraction the next time. That is why steps in a running task are cheap and a fresh start costs more.

The quick start task, changing one line in one resource, took three to four requests and about 25 credits. A task that reads a whole server folder can cost ten times that.

The extension shows $0.0000

The cost line in the chat comes from the original editor tool and counts dollars of a paid provider. SwisserAI bills in credits instead, so it always shows zero. Your real usage is on the website.

Where to see what you used

Open your account settings. Usage shows requests, tokens and credits used this month for the whole account. Under API access each key, including the one of the extension, shows its own requests and tokens.

Habits that save credits

  • Name the resource and file in the task

    The agent goes straight there instead of searching and reading many files.

  • Open only the folder you work on

    A smaller folder means a smaller file list in every request, and less to search.

  • One change per task, then Start New Task

    A long chat is sent again with every step. A fresh task starts small.

  • Work in one block of time

    Text read a few minutes ago is reused cheaply. After a long break it is read at full price again.

  • Keep big files out

    Add maps, models and logs to .swisseraiignore so they are never read.

  • Use Ask for questions

    Ask mode cannot edit, so it does not prepare edits you did not want.

  • Split very large files

    A 3,000 line client.lua is read in full again for every edit. Several smaller files are cheaper to work on.

Long chats

The bar at the top of a task shows how full the chat is: tokens used and the limit of the model (131,072 for swisserai-free, 262,144 for swisserai-premium).

When it gets close to the limit, the extension condenses the chat: the agent writes a summary of what happened so far and continues from that. The chat shows Context Condensed. Since version 0.1.1 this summary reuses the text the agent already read, which makes it several times cheaper than before.

A summary cannot keep every detail. If the agent forgets something after a condense, repeat it in one sentence. Or start a new task with a short note on where you are. The threshold can be changed in the settings under Context.