Argoforgebeta
Argoforgebeta
Kairos tells Claude Code when to run tests and update docs, stops tool calls you banned or already refused, and shows what your work costs in tokens and time.
Step by step
Three moments: the start of a session, every tool call, the end of a turn.
Once per session a short note is added: tests and docs right after each edit (the default) or once before a commit. You choose on the Kairos page.
Before the call, the app checks it on your Mac: was this exact call refused earlier in the session, and does an edit put a path from your Mac into shared code?
The server stops a call that breaks your ban, repeats a refused call, or puts your Mac's path into shared code. Claude Code sees the reason.
Only figures are sent: tokens, time, tool calls by name, errors, refusals, context compactions, the model id the turn ran on, and which Argoforge agents and skills it used.
Savings and speed for 1, 3 or 7 days on the Kairos page, with how well the turn estimates held. Argoforge mode is compared with your own flow in dollars per turn, by session, only between turns at the same model and context size. Optionally, a test: half of your sessions run without Argoforge's help, to show what it gives.
Your data
Not on the server: the guard and the reports are plain rules and counting. A ban is a rule in your brain on the server; the guard checks it before each call.
Rules
A ban covers terminal commands (build, tests, commit, push, deleting, or a command pattern) or an MCP server. It works from the next call, and the Kairos page can switch it off. In the test's sessions without Argoforge the guard is off, bans included.
If you refused a call in this session, the exact same call is stopped.
A path from your Mac in shared code is stopped once. If you really need it, the same call goes through the second time.
The guard waits for the server at most 2 seconds; if the server is slow or down, the call goes ahead.
Before each turn the server estimates its cost from your turns most like it — context and prompt length, the model, and whether the session's previous turn called subagents — as a range from the 25th to the 75th percentile. About half of the turns should land inside it.
A server experiment stops itself, within minutes of the evidence, when a variant measurably costs more per turn, errs more, or splits the work into more turns per session.
A tool result over 10,000 bytes counts as heavy in the reports.
How to use
Keep “right away” when you edit and check often; pick “before commit” to test a long run of edits at once. Tests you ask for always run at once.
Numbers over a week are steadier than a single day.
Each ban has its own switch on the Kairos page.
The page shows measured turn figures only. v1's whole-session estimates are not shown: the server never sees session transcripts — they stay on your Mac.
Saying “never …” does not create a ban in v2 yet: your Claude can draft one, but nothing saves it to your brain on the server. Bans already in your brain work, and each has a switch on the Kairos page.