Spend fewer credits

How do I use fewer AI credits in Mel? Use your own provider key (0 Mel credits), pick the Cost-efficient Auto mode, keep each conversation to one task, and run plain shell commands yourself instead of asking the agent to. This guide covers each of these, and how to check what a turn cost.

How a turn is billed

1 credit is 1 cent. A turn costs what the model that answered actually charged, plus a 20% margin. Two things decide most of that:

• Which model answered. Models differ a lot in price.
• How much the model read. On every step the agent sends the conversation so far, so a long conversation costs more per step. Providers charge less for context they've already seen, and so does Mel, but a shorter conversation is still a smaller bill.

So the tips below either pick a less costly model, or send less.

1. Use your own provider key

A turn on your own key costs 0 Mel credits; you pay your provider directly. Groq and Google AI Studio give free keys.

1. Open Settings ▸ AI ▸ Provider API keys.
2. Paste the key into its provider's row and save.

See Get a free model key. Web search, command suggestions and voice input still count against your plan's daily limits, because they run on Mel's own accounts.

2. Pick the Cost-efficient Auto mode

Open the model chip (or type /model) and choose Cost-efficient. It prefers the least costly model that can do the step, and moves up only when it has to. The mode you pick becomes the default for new conversations. See Models & pricing.

If a model offers a Reasoning level, a lower level thinks for less time and uses fewer tokens. See Models & routing.

3. One task per conversation

A conversation carries everything said in it, and the agent re-reads it on every step.

• Start a fresh conversation (/mel, or Ctrl+Enter, ⌘Enter on a Mac) for a new task instead of continuing an old one.
• When a long conversation still has work left, type /compact, or /compact <what to keep>, to fold the older turns into a short summary. Mel also does this by itself when a conversation nears the model's limit. See Conversation management.
• Keep facts the agent needs every time in project memory or MEL.md, so you don't paste them into each conversation. Keep MEL.md short: it's read on every turn. See Rules and MEL.md.

4. Give it exactly what it needs

• Attach the output you mean. Ctrl+↑ (⌃↑ on a Mac) attaches the last command and its output to your message, so the agent doesn't have to run it again or search for it.
• Name the files. "Fix the null check in src/parse.rs" costs fewer steps than "find the bug".
• Plan big jobs first. For a large change, /plan <task> lets you fix the approach before any code is written, instead of paying for work you then undo. See Planning.

5. Do the free parts yourself

• Shell commands are free. Run git status or npm test yourself instead of asking the agent to. In an agent conversation, start a line with ! to run it in the shell.
• CLI coding agents are free in Mel. Claude Code, Codex and others run in a Mel tab on their own accounts and never use Mel credits. See Embedded CLI agents.

Check what you spent

• Per answer: the usage line under each answer shows the model, the tokens and, when Mel measured it, what it cost in credits. Expand it to see why that model was picked. The conversation's running total is there too.
• Per day: Billing and Usage ▸ Usage history shows the last 30 days, day by day and per conversation.
• Running low: a small banner warns you when your monthly credits are low, with a rough count of turns left.

See Billing & usage.

FAQ

What is the cheapest way to use Mel's agent? Your own provider key: turns on it cost 0 Mel credits. Free keys from Groq or Google AI Studio work on the Free plan.

Does a long conversation cost more? Each step re-reads the conversation, so yes, a longer one costs more per step. Start fresh for a new task, or use /compact.

Does Cost-efficient mode give worse answers? It still moves up to a stronger model when a step needs it. It just starts low.

Do shell commands or Claude Code use credits? No. Shell commands and CLI coding agents never use Mel credits.