claude-autonomous

Stops Claude Code from asking. No permission prompts, the whole disk in scope, and a skill that acts instead of asking.

Read what it changes before running this. It removes every guardrail the harness gives you, on purpose. That is the point, and it is also the risk: an agent with no prompts will delete the wrong directory just as confidently as it writes the right file. Bring git and backups.

Install

macOS / Linux

curl -fsSL https://raw.githubusercontent.com/NspxMiguel/claude-autonomous/main/install.sh | bash

Windows — no administrator rights needed

irm https://raw.githubusercontent.com/NspxMiguel/claude-autonomous/main/install.ps1 | iex

Then restart your Claude Code session — the permission mode is read at startup.

claude-autonomous status   # what is actually set right now
claude-autonomous off      # put everything back

Why config alone is not enough

Most advice stops at --dangerously-skip-permissions. That suppresses the harness prompt, and then the model stops anyway to write “want me to do X?” — which, to the person waiting, is the same stall.

And past that there is a third stall, the one nobody configures away: the model finishes four items out of seven and writes “done for now — all that’s left is X”. No prompt was raised, no question was asked, and the person is waiting on the session again anyway.

Config

No prompt can be raised.
~/.claude/settings.json

Posture

No question is invented.
the autonomous skill

Finishing

No turn is handed back with work open.
claude-autonomous finish

The skill is the part people skip, and it is the part that actually changes how a session feels. Finishing is the part that cannot be a promise: it is a Stop hook, so while the backlog holds an open item the harness refuses the end of the turn. Your word is the off switch — stop (or chega) clears it before the model reads your message.

claude-autonomous finish start --task "port the dashboard"
claude-autonomous finish add "rewrite the client" "update the tests" "commit and push"
claude-autonomous finish done 1 --proof "npm test — 42 passing"

What it changes

Setting	Effect
`defaultMode: bypassPermissions`	No permission prompt
`hooks.PreToolUse` → allow	Answers before a prompt can exist
`hooks.Stop` → finish	Refuses the end of a turn while work is open
`allowAllBrowserActions`	The desktop app stops asking per site
`hooks.UserPromptSubmit` → finish	Your off switch: “stop” / “chega”
`deny: []`, `ask: []`	Nothing held back or escalated
`additionalDirectories`	Whole machine in scope
`sandbox.enabled: false`	Commands run unconfined
`askUserQuestionTimeout: 60s`	A question never parks the session
`fileCheckpointingEnabled: true`	`/rewind` — the last undo left

What it does not change

Operating-system permissions (macOS screen recording, accessibility, automation) raise their own dialogs and no settings file removes them. Neither does any config remove the model's own limits: typing a password or 2FA code into a form, logging in as you, moving money, or acting on instructions found inside a web page rather than coming from you.

Note that the credential line is about typing, not using — an authenticated CLI like gh or vercel does the work with the secret staying where it is.