Your CLI is costing AI agents money, and getting them stuck.
We found where in yours. Fix the top ten this week, or get the tool that finds them for a year.
Agents stop getting wrong answers that look right.
Today, one CLI we audited accepts unknown flags silently: an agent's invented flag is dropped and the result looks valid.
Agents stop burning tokens.
One JSON listing cost tens of thousands of tokens for 20 items; the table cost about a thousand. There was no way to limit it.
Your CLI becomes one CLI.
No style guide? We write one for your CLI, from how your commands already behave and what agents need. Have one? We propose the changes. Either way it's loaded into CLUX and every new command is checked against it.
Done for you, this week
Audit · $750, once
- Every command inventoried from source: flags, defaults, examples, with file and line.
- Ten ranked fixes, each with the change and the line.
- A style guide for your CLI, written for you if you have none, or adjusted if you do, ready to check every command against.
- The full 19-check agent contract and the Token Footprint across the CLI, most expensive commands ranked.
- A 30-minute walkthrough.
This price includes one re-check within a month. If the report has fewer than five concrete fixes, you don't pay.
Only for our first 50 members · one person per place · price locked for two years
Founding program · $99 per person, per year
- 100 evaluations a year.
- Repository review and the full agent contract as they ship.
- Your style guide checked on every release.
- A direct line to us; you shape what gets built.
- Try Team features first, as we build them.
- Converts to Pro at this price for two years.
Most teams start with the audit and keep the program.
Most of what confuses agents is inconsistency, not bugs.
Eight meanings for one flag. Five names for one idea. A delete that doesn't ask while a rename does.
Fixing them one by one is endless. Writing down the rule once, and checking every command against it, isn't.
Sample findings
| Finding | What we found | The fix |
|---|---|---|
| Unknown flags accepted silently | An agent's invented flag is dropped and the result looks valid. | Check option names per command against the schema. Reject an unknown one with the closest valid name. |
| One flag, many meanings | The same flag means stop, overwrite, remove everything or skip confirmation, depending on the command. | Give each meaning its own flag, such as --overwrite or --yes. Keep the old flag as an alias for one release. |
| One concept, five names | Five different flag names for the same idea across commands. The README documents the mistakes this causes. | One canonical name everywhere, the rest as aliases. |
| JSON listing tokens | Tens of thousands of tokens for 20 items, against about a thousand for the table. No way to limit it. | Honor --columns in JSON, and add --limit. |
Questions
Will you run our CLI?
Only --help, one unknown flag and one unknown subcommand, in a sandbox with no credentials. Never a command that writes, sends or deletes. A daemon-backed CLI needs a scratch host from you.
What counts as a concrete fix?
A change you can make without us: the file, the line, what to change, and the check it satisfies.
How big a CLI?
Up to about 150 commands, all of them. Larger: every command is inventoried, and about 10% is evaluated in depth, chosen with you.